open-standards/wai/CHANGELOG.md

# WAI — changelog

Notable changes to the WAI standard and the reference implementation.
Newest at the top. WAI is a draft standard; until v1.0 freezes, the
wire format may change and such changes are recorded here.

## v0.4.0 — current

### `wai.live.ads`: a room's cues as server-guided ad breaks, and the break form of `wai.video.ssai`

- **The extension** ([live-ads](extensions/live-ads.md)) registers the record `wai.live.ads`.
  The room authority signs an **ad policy**, `WAP1`: the grant the ad decider runs under, whose
  funds ceiling is the room's ad budget; the slots of a break; the avail of a break opened with
  no planned duration; a preload window; skip and jump restrictions, a skip control and
  snapping; and the bases of the asset-list and creative URIs.
- **Breaks from cues.** A room's breaks are its signed `cue` events of `wai.live.events`,
  unchanged: an out and its in, or a break of known duration; a chapter opens nothing. A
  post-roll is a break its author annotates with a `WAB1` `meta` event right after it, and ends
  the program. A verifier refuses an out or a break inside another break, an in with no open
  out of its `cue_id` and author, a break of no duration, an out with no planned duration when
  the policy has no open avail, an in whose stated duration is not the one that passed, an
  annotation that does not read or that follows anything but a break, and any cue after the
  post-roll; the room refuses a cue by a principal whose grant does not permit `cue`.
- **Shared manifests.** Each break is one Interstitial `EXT-X-DATERANGE`
  (`draft-pantos-hls-rfc8216bis-22` Appendix D) with `CUE`, its duration or planned duration,
  `X-ASSET-LIST`, `X-RESUME-OFFSET`, `X-PLAYOUT-LIMIT`, `X-SNAP`, `X-RESTRICT` and
  `X-SKIP-CONTROL-OFFSET`, a closing tag at its in, and a Preload date range (Appendix F); and
  one DASH alternative-presentation event of ISO/IEC 23009-1:2026, a replacement for a mid-roll
  and an insertion for a post-roll, read from the sixth edition's schema and its examples, with
  the attributes whose meaning that revision could not state left unwritten. Every value is a
  function of the signed cue, its annotation, the policy and `t0`, and the `ID` is derived from
  the cue's digest, so every edge writes the same bytes.
- **Asset lists.** The one URI a break names is where each viewer is decided: an HLS asset list
  or a DASH list MPD, written exactly, each creative's URI naming it by its hash, filled by
  `wai.video.ssai`'s decision over the viewer's pod; a viewer who joins a break running is
  filled from what remains of it. An empty fill is `{"ASSETS":[]}` in HLS and no list MPD in
  DASH.
- **`wai.video.ssai` gains form 2, the break form.** A receipt names the room, the break by its
  cue's digest, the session, the viewer's offset, the asset list's SHA-256 and format, the
  avail, the slots and the funds the decision ran with, beside the stitched form's pod hash,
  impressions, Merkle root, billable spend and energy figure. It is carried as JSON, signs
  under its own domain (`wai:video-ssai-break-receipt`) and shares the identity domain, so the
  figure profile `wai:video-ssai-receipt-id` names either form. Every stitched receipt keeps
  its bytes, and its legacy pin is unchanged.
- **Verification end to end**: the room, the policy, the schedule, the renderings, then each
  receipt: its own rules, the decider's key and grant reference, the break, the decider grant's
  window and revocation, the pod hash, the decision recomputed, the asset list's digest, one
  receipt per session and break, and the spend of every receipt together within the decider
  grant's ceiling (prices in millionths of a minor unit of its currency).
- **Twenty refusal reasons**, from `ad_policy_malformed` to `funds_exceeded`.
- **Reference and corpus.** `wai-rs` feature `live_ads` (`src/live_ads.rs`), and
  `SsaiBreakReceipt` in `src/wai_video_ssai.rs`. `live-conformance/ads/`, written from fixed
  seeds by `wai_live_ads_vectors`: a live break with a viewer at its start and one who joins it
  running, a mid-roll under restrictions with an empty fill, a program closed by a post-roll,
  chapters with a planned out, a gap, and one refused room per reason; each accepted room's
  HLS tags a file of its own. Checked by `tests/live_ads.rs` and by
  `live-conformance/live_ads_verify.py`, a second implementation that shares no code with
  `wai-rs`, verifies each room with `live_rooms_verify.py`, writes every tag, event and asset
  list itself, and requires every mutated copy of the corpus to fail. The registry lists the
  record and the magics `WAP1` and `WAB1`.

### `wai.live.watermark`: forensic watermark sessions, matched and disclosed one at a time

- **The extension** ([forensic-watermark](extensions/forensic-watermark.md)) registers the
  record `wai.live.watermark`, for a room's Lane A delivery (SPEC §6.5). The edge that chooses
  each segment's A or B variant for a viewer, as ETSI TS 104 002 has an edge choose it, signs a
  **session record**, `WWS1`, when the session ends: the room and the content (a slot or a
  publisher, and its media kind), the span, the KIDs of the content keys the variants were
  encrypted under, the TS's watermark token by its SHA-256, the token's vendor and pattern
  length, how the edge had the pattern (in the clear, decrypted, or derived by the vendor's
  core), a bitmap of the positions it sequenced and the variant at each, its counts of
  sequenced and unmarked segments, a salted commitment to the session's grant and principal,
  where it keeps the record, and when it is deleted.
- **The token** is the TS's CWT, read as the TS and RFC 8392, RFC 8949 §4.2, RFC 9052 and RFC
  9053 have it: deterministically encoded CBOR, a COSE_Mac0 under HMAC 256/256 or a COSE_Sign1
  under ES256, optionally in the CWT tag, and the claims `wmver`, `wmvnd`, `wmpatlen`,
  `wmsegduration`, `wmpattern` (in the clear, or a COSE_Encrypt0 or COSE_Encrypt) or `wmid`,
  `wmopid` and `wmkeyver`, with `exp` and `iat`. A form the RFCs forbid is `token_malformed`,
  and one they allow that this revision does not read (a floating-point date, an untagged or
  nested message, another algorithm) `token_unsupported`. The record names the token by digest
  and does not carry it: a token can carry `sub` or a vendor's `wmid`, and can be presented
  again until it expires. The token's MAC or signature is the edge's to check.
- **No viewer in the record.** It names no grant: every participant holds a room's grants, and
  a grant names its principal. Only the room authority, to which the edge gives each record's
  opening, can link a session to a principal. Records, tokens and openings are kept within the
  room's region pin and its retention bound.
- **Matching.** A claimant signs the variant sequence it recovered from a copy, `WWQ1`, with
  its erasures. A session agrees when it sequenced every stated position with the stated
  variant; a match is the one agreeing session among the records given (`no_match`,
  `ambiguous_match` otherwise). The TS leaves detection's statistics to its vendors, so the
  rule has no tolerance and no threshold, and states how many positions were stated.
- **Disclosure.** For one matched session the room authority signs a `WWD1` that opens its
  commitment; a third party checks it against that record, the sequence and the grant, and
  learns nothing of other sessions.
- **Lane A only.** A record names at least one content key (`not_protected`), and a room with
  a confidentiality record has no edge that can mark it (`e2ee_room`): its objects' AEAD binds
  each to its track, group and object.
- **Refusal reasons**, from `record_invalid` to `disclosure_expired`. jwp-receipts §7.5
  registers the key-log authority `live/watermark` for an edge.
- **Reference and corpus.** `wai-rs` feature `live_watermark` (`src/live_watermark.rs`), with a
  deterministic-CBOR reader of its own and no new dependency. `live-conformance/watermark/`,
  written from fixed seeds by `wai_live_watermark_vectors`: tokens read and refused, an
  edge's records of four sessions with a token in each mode, sequences matched, unmatched and
  refused, disclosures accepted and refused, one control for each reason. Checked by
  `tests/live_watermark.rs` and by `live-conformance/live_watermark_verify.py`, a second
  implementation that shares no code with `wai-rs`, requires every mutated copy of the corpus
  to fail, and requires the corpus to catch a verifier that leaves out any one of its checks.
  The registry lists the record and the magics `WWS1`, `WWQ1` and `WWD1`.

### Live L7b: the DVR window, the seek power, and WMH1 version 2

Additive: no existing record changes its bytes, and every existing corpus vector keeps its
records.

- **live-recording §8** gains what a relay that keeps a **DVR window** of a room's tracks
  does with it (§8.2 to §8.5): a window by age (at most 24 hours) and by bytes, evicted
  oldest first by whole group, each media group with its receipts group; a FETCH answered from
  the cache and the window together, groups in the requested order; what was evicted marked
  with an End of Unknown Range (MOQT §11.4.1.2, `0x10C`), never with an unmarked gap or an End
  of Non-Existent Range, since an evicted object existed; an object that cannot be served
  inside a group marked from the group's start, since -22's End of Range moves to a later
  group; a harvest window whose end lies ahead as a SUBSCRIBE with that range and a fill;
  decisions read from the room's events tracks withholding what they deny or redact, never
  replacing it; every object checked again against its receipt after a restart. §8.6 maps the
  window onto DASH's `timeShiftBufferDepth` and an HLS live playlist. §8.1 is the take
  addressing as before. §8.2 also says what a relay never keeps: a room with a confidentiality
  record, known by its first sealed object. The relay keeps nothing of the room from then on,
  deletes what it kept of it before, and remembers the room across a restart. confidentiality
  §9 points to it.
- **live-rooms**: the **seek** power, bit 4 of `powers` (§3.3), and **§10.6 Seeking back**:
  a FETCH, or a fill, that starts below a track's live edge is served only to a grant with the
  power, and refused `UNAUTHORIZED`, `seek_not_granted`, otherwise. Roles and grants may now
  set bits 0 to 4. `wai-rs` `live_room::seek_check`; the corpus gains a `dvr` room, three
  refusals (a grant's seek beyond its role, power bit 5 in a grant and in a role) and a
  `seek` section of requests served and refused, which `live_rooms_verify.py` checks too.
- **live-metering**: `WMH1` **version 2** (§3.9) adds a hop's DVR figures before `sig`: the
  objects and bytes served from its window, those it wrote to it, the bytes it held in
  byte-seconds, and the objects it dropped because they did not verify; with their rules
  (§3.6) and bound `B₂(M) = 600 + 180 × (M + 2)` (§3.7). A hop with no window signs version 1.
  The receipt stays version 1 and does not sum the figures. The corpus gains a version 2 relay
  record and four refusals; `live_metering_verify.py` reads both versions.

### wai-quantum 0.6.9: cut-rank amplitudes with the circuit's own global phase

Additive. Every pinned digest is unchanged.

- **`PhasedTableau`** (`quantum_tableau`) is a stabilizer tableau with its global phase.
  - It tracks one basis state `r` of the support and its exact amplitude
    `⟨r|ψ⟩ = 2^{-k/2} e^{iπm/4}` through every gate.
  - Paulis, phase gates, CX, CY, CZ and SWAP move `r` and the phase directly.
  - For a Hadamard on qubit `a`, the amplitude of `r ⊕ e_a` comes from the stabilizer
    whose X part is `e_a`, found by reducing the stabilizers (about `n³/64` word
    operations). `√X` is `H S H`, exactly.
  - Tested against the dense state vector after every gate of 40 random circuits
    (2 to 8 qubits, over 2,000 gates). A mutant of the Hadamard update is caught.
- **`ProductMeasured::with_last`** keeps the graph form, order and steps with other closing
  gates. **`ProductMeasured::scale`** multiplies every amplitude by a factor, folded into
  one qubit's weights. Tested against fresh builds.
- **The service's `cutrank` method** returns amplitudes with the circuit's own global
  phase.
  - The factor is `e^{iΘ} a/b`. Here `a` is the tracked amplitude of the Clifford part,
    `b` is the graph form's amplitude at the same basis state (closing gates set to the
    identity), and `e^{iΘ}` is the phase the lowering to Clifford gates drops. That
    phase is read gate by gate from each gate's matrix against its lowering's.
  - Past about a minute of tracking (counted from the Hadamards), the result says
    `global_phase` is not fixed, as before.
  - `auto` now tries `cutrank` for amplitudes too, before `tensor_network`.
  - Tests compare amplitudes with the state vector's directly (no phase aligned). The
    circuits include phase-carrying lowerings: `rz`, `rx`, `ry`, `u`, `p`, `cp`, `rzz`,
    `rxx`, `iswap`, `sx`, `sxdg`, `cy`, `swap`, `crz` and `cu`. Two mutants are caught:
    the lowering phase dropped, and the tracked amplitude ignored.

### wai-quantum 0.6.8: threads only where they pay, and a cut-rank search that recounts only what moved

Additive. No result changes a bit: every kernel path gives the same bits at any thread count,
and every pinned digest is unchanged.

- **State-vector passes** (`quantum_sv`): a gate, a slice pass or a fixed-chunk sum now spawns
  at most one thread per `2¹⁸` amplitudes. Each pass spawns its threads afresh, and below
  that size a thread costs more than its share. On a shared 32-core host, a gate from 14 to
  18 qubits ran 5 to 14 times slower with a thread per core than on one thread. A count that a
  test forces is still taken as it is, so the split paths stay tested at small sizes. The
  ignored `gate_timings` test prints the times.
- **The cut-rank search** (`ProductMeasured::anneal_order`) moves one qubit at a time. Seven
  moves in eight go at most four places, and the rest anywhere. Only the cuts between the
  two places are recounted, so a move costs about a ninth of a full recount. On the 70-qubit
  graph-state circuit, 300 000 moves reach a 29-bit widest cut in about six seconds. In
  tests, every move's recounted cuts are checked against a full recount.
- **The service's `cutrank` method:**
  - The planner makes one short search, enough to bound the job's memory and work.
  - The run searches up to 300 000 moves from each of four seeds, on threads of their own,
    with the planner's order among the candidates.
  - It keeps the cheapest order whose widest cut is no wider than the planner's, so a job
    stays inside the memory it was admitted with.

### wai-quantum 0.6.7: tensor networks that split controlled gates and fix projections

Additive. `Network::amplitude` and `Network::simplify` are unchanged, so every pinned digest is too.

- **`Network::amplitude_split`** (`quantum_tn`) splits each gate on two or more qubits at
  every qubit it is block-diagonal on: a control, or a qubit it is diagonal on. There
  `G = Σ_a |a⟩⟨a| ⊗ G_a`, so that qubit gets the delta `δ(out, in, k)` on its wire, and the
  blocks `G_k` keep the rest, joined by a bond of dimension 2. Only zeros of the matrix are
  read. A cut through a controlled-Z or controlled-X then costs 2, not 4.
- **`Network::reduce`** fixes every index that some tensor allows only one value of: each
  tensor holding it keeps its entries at that value and drops the index, until nothing
  changes; then it simplifies. Only zeros are read and entries selected. Input and output
  projections run through diagonal gates and split deltas, so a controlled-Z beside a
  projected output becomes scalars. Open indices are never fixed.
- On 61-qubit controlled-Z circuits of depth 8, 12, 16 and 20, the search (64 trials) finds
  single-amplitude trees of log₁₀ 5.35, 8.25, 11.22 and 14.26 multiply-adds on split, reduced
  networks, against 8.03, 12.46, 16.09 and 19.64 on whole-gate, simplified ones.
- `examples/tn_cost`, the CLI's `wai-quantum tn` (`--whole-gates` for the old networks) and the
  service's `tensor_network` method build split, reduced networks.
- Tests: split networks, reduced or not, and reduced whole-gate networks give the state
  vector's amplitudes on circuits of controlled-Z, controlled-X both ways, a controlled
  random unitary, random diagonal gates, Toffoli, CCZ and random two-qubit unitaries, with
  outputs projected and open. CZ and CX split and a random unitary does not; CX splits at its
  control. On a 4 × 5 grid of controlled-Z, the split tree costs under an eighth of the
  whole-gate one. Three mutants are caught: the delta's entry, the block-diagonal test, and
  the value kept by `reduce`.

### wai-quantum 0.6.6: Clifford circuits measured in a product basis, exactly, by cut-rank

Additive. Every pinned digest is unchanged.

- **`Tableau::graph_form`** (`quantum_tableau`) writes a stabilizer state as local Cliffords
  (`H^h S^s Z^z` per qubit) on a graph state. Tested by rebuilding the state from the form
  over 40 random circuits.
- **`quantum_cutrank`** (a new feature, in `full`) sums a Clifford circuit's state measured
  after one closing layer of single-qubit gates at any angles:
  `⟨x| ⊗ V_q |ψ⟩ = 2^{-n/2} Σ_z ∏_q W_q[x_q][z_q] (−1)^{Σ_{a~b} z_a z_b}`.
  - The sum goes one qubit at a time along an order, in a table of `2^{cut-rank}` entries.
    Each qubit is an in-place pass: a merge where the boundary space loses a dimension, the
    qubit's signed terms, and a doubling or an XOR-shifted sum for its edges ahead.
  - The passes depend on the graph and the order alone, so they are planned once per state.
  - By Parseval, the table's norm is a prefix's probability, so the same passes sample
    exactly: `amplitude`, `probability`, `sample`.
  - `Precision::Single` stores the table in `f32` with `f64` arithmetic: half the memory and
    traffic. `anneal_order` searches orders for less work (`work` = Σ 2^cut) and never
    returns a worse one.
  - Every reduction is summed in fixed chunks, so a result is the same bits at any thread
    count.
  - Tests:
    - amplitudes equal the state vector's in probability and relative phase (30 circuits up
      to 12 qubits);
    - forward, reverse and shuffled orders agree (24 to 54 qubits);
    - single-precision tables agree with double;
    - samples follow the exact law and are seeded;
    - annealing lowers the work and keeps the amplitudes;
    - cut-ranks are the entanglement of GHZ and product states.
  - Five mutants are caught: the merge's signs, the sampling choice, the doubling's sign,
    the fold's coefficient, and the graph form without its S.
- **Examples:**
  - `cutrank_qasm` reads an OpenQASM 2 Clifford bulk with a closing single-qubit layer, and
    prints cut-ranks, annealed orders, amplitudes and exact samples. On a 70-qubit
    graph-state sampling circuit, an annealed order keeps every cut at 29 bits or fewer: one
    amplitude in about half a second, one exact sample in 1.1 to 1.5 s, with a largest
    table of 4 GiB.
  - `patch_xeb` scores measured shots of a patched circuit by patched cross-entropy, each
    patch simulated exactly on the dense state vector.
  - `tn_cost` prints the contraction cost of one amplitude as `find_path` finds it.

### wai-quantum 0.6.5 and the service: programs that measure in the middle

Additive. Every pinned digest is unchanged, and a program that measures only at the end
runs exactly as before.

- **`Tableau::measure`** (`quantum_tableau`) measures a qubit of the wide tableau and leaves
  the state the outcome selects. A random outcome is the caller's to choose, so a branch
  can be followed on purpose. The anticommuting stabilizer is multiplied into every other
  row at once, each row's phase summed mod 4 in two bit planes. `Tableau::is_random` says
  in one column scan whether an outcome is random.
  - Tested against the state vector's projection, Pauli string by Pauli string, over 60
    random circuits of six measurements each.
  - On wide tableaux, measuring every qubit lands inside the support held before, with one
    random outcome per dimension.
  - Three mutants are caught: signs kept, the Y phase swapped, the destabilizer not moved.
- **`StateVector::probability_one` and `StateVector::collapse`** (`quantum_sv`): a qubit's
  probability of reading 1, summed in fixed chunks, and the state a reading leaves.
- **The service runs dynamic programs.** It reads measurement and `reset` anywhere,
  `if (c[0]) …`, `if (!c[0]) …`, `if (c == 3) { … } else { … }`, OpenQASM 2's
  `if(c==n) gate;`, and `bit b = measure q;`.
  - Such a program runs on `statevector` or `stabilizer`. Its shots split at each
    measurement and reset as independent shots would, and each outcome runs on in its own
    copy of the state.
  - Expectation values are the exact mixture over the branches, of the state before the
    final measurements.
  - Tests: teleportation gives ⟨Z⟩ = cos θ and ⟨X⟩ = sin θ exactly. It agrees between the
    tableau and the state vector, reset and re-measurement behave, and a bit-flip syndrome
    corrects the flip. Four mutants are caught.
  - Loops, noise in such a program, and amplitudes of one are refused with their reasons.

### `wai.live.moderation`: review policies, signed decisions, and statements of reasons

- **The extension** ([live-moderation](extensions/live-moderation.md)) registers the record
  `wai.live.moderation`. The room authority signs a **review policy**, `WMP1`, chained to the
  one it replaces and in effect from a room-clock instant: which event kinds and media are held
  for review before delivery (pre) or reviewed after it (post), the reviewers, each one's
  deadline and fallback (allow, deny or hold), the trusted flaggers it recognises, the reason
  codes with their ground, provision or clause and database category, the territory, and the
  redress text. Every policy keeps the room descriptor's review field (live-rooms §3.6).
- **Decisions.** A moderator's verdict is a `decision` event of `wai.live.events`, unchanged,
  followed in its chain by a `meta` event carrying the **statement**, `WMS1`: its target (an
  event, or a media object such as a recorded chunk named by its leaf), the target's `prev`,
  the decision it supersedes, its basis (own initiative, a notice, a complaint, the policy's
  fallback), its use of automated means in detection and in the decision, and its ground,
  explanation and facts. Decisions on one target form a line by digest; an appeal's outcome
  supersedes the decision complained of. The room's clock signs each fallback the policy
  applies, at or after the deadline. A decision by a principal without the moderate power is
  refused by the room verifier (`kind_not_permitted`).
- **Notices and complaints**, `WMN1`, are signed by whoever submits them: a notice names an
  item, its ground and the submitter's statement and contact (Article 16(2)); a complaint
  names a decision (Article 20). The contact is never exported.
- **Review and redaction.** Each item's state over time (held, pending, allowed, denied,
  redacted, or the fallback at its deadline), what a relay holds and delivers, and what a sink
  shows. A restricted item is withheld, never replaced: a gap it leaves in its author's chain
  is explained by the moderators' signed statements, walking the chain back through each
  withheld event's `prev`; a clip over a redacted chunk keeps its range proof, withholds the
  chunk's bytes (`redaction_not_applied` when they are given), and presents the chunk again
  once a later decision restores it. A recording package under moderation is specified only.
- **The statement-of-reasons export.** Each restricting decision maps onto one record of the
  DSA Transparency Database's statement-of-reasons format (the schema in force from
  1 July 2025, which its documentation calls API v2), written byte for byte: every member,
  its source in the WAI records, closed lists and length limits, the sixteen categories, the
  content types, and every Article 17 element with WAI's field for it. It carries no
  principal key, notice contact or content; `source_identity` is never written. A record that
  cannot be written is refused `sor_incomplete`, naming the member.
- **Twenty refusal reasons**, from `policy_malformed` to `redaction_not_applied`.
- **Reference and corpus.** `wai-rs` feature `live_moderation` (`src/live_moderation.rs`);
  live-recording's clip verifier gains `verify_clip_exempting`, which `verify_clip` calls
  exempting nothing, so its corpus is unchanged. `live-conformance/moderation/`, written from
  fixed seeds by `wai_live_moderation_vectors`: a pre-moderated chat room whole and as a
  subscriber receives it, a notice and a trusted flagger's notice with their decisions, an
  appeal that reverses a decision, live-recording's lesson session with a chunk redacted and
  then restored, one refused room per reason, an export refused, and clips of the redacted
  take; each export a file of its own. Checked by `tests/live_moderation.rs` and by
  `live-conformance/live_moderation_verify.py`, a second implementation that shares no code
  with `wai-rs`, verifies each room and clip with its corpora's second implementations, writes
  every export itself, and requires every mutated copy of the corpus to fail. The registry
  lists the record and the magics `WMP1`, `WMS1` and `WMN1`.

### wai-quantum-service: every job's energy, signed with its result

- **What a result carries.** `energy` holds the CPU energy of the job's worker process, from
  its start to its exit, inside the body the receipt seals. It is given as `joules_micro`,
  with the acquisition class `OnChipCounter`, a declared uncertainty
  (`relative_ppm`, `window_us`) and the counter's filtering state, in
  [energy-measurement](extensions/energy-measurement.md) §2's names. It also states its
  coverage and what it excludes.
- **How it is read.** On macOS, the service reads the worker's kernel count
  (`proc_pid_rusage`, `ri_energy_nj`, plus energy system services billed to it) after the
  worker has exited and before it is reaped, waiting with `waitid(WNOWAIT)`. An exited
  process has left every core, so the count is complete. The declaration covers the
  rounding to whole microjoules, and meets the §2.3 floor at every size.
- **What it does not claim.** A `statevector_single` job runs on the GPU, which the counter
  does not read, so its figure is `partial`. A job the service cannot read is
  `{"unmetered": why}`, never zero joules.
- **Agents.** The MCP summary states the figure and its coverage.

### `wai.live.steering`: each client's pathway, chosen by the joules its hops signed

- **The extension** ([energy-steering](extensions/energy-steering.md)) registers the record
  `wai.live.steering`. The operator signs a **steering policy**, `WSP1`: the pathways (each a
  pathway ID of the content description or a clone of one, through another host), each
  pathway's hops (edge or interior, a hop in one pathway only), the quality floor, the latency
  ceiling, each pathway's capacity and minimum share, the hysteresis margin, the TTLs, and the
  period and window over which the hops' per-hop records (`WMH1`) are read. The new key-log
  authority `live/steer` lets an operator key sign policies and decisions.
- **The evidence.** A pathway's figure is its hops' joules over the window, per media byte its
  edge hops delivered or per microsecond of delivery at the floor tier or above, with the
  records' common acquisition class. The window ends on the last period past a lag, and every
  hop's records must tile it exactly, so every pathway is read over the same span and a
  record left out makes the pathway **stale**, never cheaper. A pathway with an unmetered hop,
  hops of two classes, stale records or no delivery has **no figure**: a decision that gives
  one a figure is refused `unmetered_as_zero` or `classes_summed`.
- **The rule.** Pathways are ordered by class (calibrated instrument, counter, model,
  estimate), within a class by figure, compared by exact products, and then by ID; a pathway
  with no figure comes after every pathway with one, so not measuring is never rewarded. A
  client with a short buffer or at its quality floor is held; a session's draw, fixed for the
  session, places it on a pathway by that pathway's minimum share whatever its evidence; a
  client moves off its pathway only for a figure lower by more than the margin. The output is
  the priority, the TTL and the reason the first pathway is first.
- **The steering decision**, `WSD1`, names the policy, the records by digest, the client's
  report (read from `_HLS_pathway` and `_HLS_throughput`, `_DASH_pathway` and
  `_DASH_throughput`, and CMCD's `bl` and `mtp` in the query or the `CMCD-Request` header), the
  server's session and loads, each pathway's evidence and limits, and the output; a verifier
  recomputes all of it; ten refusal reasons.
- **The mappings.** The steering manifest of `draft-pantos-content-steering-04`, which HLS
  (`draft-pantos-hls-rfc8216bis-22` §7) and DASH (ETSI TS 103 998 V1.1.1) read, written byte
  for byte with `PATHWAY-CLONES`, a `RELOAD-URI` carrying the session, and one member,
  `science.transaction.wai-decision`, naming the decision's URI, which players ignore; the
  `EXT-X-CONTENT-STEERING` tag and the DASH `ContentSteering` element.
- **Reference and corpus.** `wai-rs` feature `energy_steering` (`src/energy_steering.rs`, with
  an RFC 9651 parser of its own); `live-conformance/steering/`, written from fixed seeds by
  `wai_energy_steering_vectors`, its figures synthetic, and checked by
  `tests/energy_steering.rs` and by `live-conformance/energy_steering_verify.py`, a second
  implementation that shares no code with `wai-rs`, reads the room and the per-hop records
  with their corpora's second implementations, recomputes every report, figure, decision and
  manifest, and requires every mutated copy of the corpus to fail. The registry lists the
  record and the magics `WSP1` and `WSD1`; jwp-receipts §7.5 lists `live/steer`.

### Live ingest and egress: WHIP and WHEP onto a room's tracks, E-RTMP v2 and SRT mapped

- **The extension** ([live-ingest](extensions/live-ingest.md)) binds the contribution and
  playback protocols of standard encoders and players to the tracks of a room. No record
  and no capability of its own; no new transport and no new token format.
- **Admission (§3).** A WHIP or WHEP bearer token is a room grant, `WRG1` in base64url. A
  gateway checks it against the room's descriptor as live-rooms §4.3 does, then its window,
  the room's revocations, its origins and its single-use nonce, and answers `401` with
  RFC 6750's challenge; a good grant without the power a request needs (publish for each
  media kind of a WHIP offer, subscribe up to the layer for WHEP) is `403`
  `insufficient_scope`. Every request with a grant carries a proof of possession of its key
  (live-rooms §10.5.3): `WAI-Room-Proof` over a single-use `WAI-Room-Challenge`, refused with
  live-rooms' `pop_*` reasons.
- **Tracks and objects (§4, §5).** A WHIP session publishes a layer-0 media track per kind,
  each with its receipts track, and the owner's catalog with a receipts track of its own.
  Each frame is one object, one `WAI1` envelope in its capability's framing: an H.264 access
  unit in Annex B, a VP8 or VP9 frame in one-frame IVF, an AV1 temporal unit in SPEC §5's
  framing, one Opus packet. The objects of a live track, from a group's first, are one
  stream of the codec. Video groups begin at keyframes; audio groups are fixed windows on
  the room clock. A track's first group ID is its first frame's room-clock time, so a
  publisher that reconnects continues above the IDs it used.
- **Time (§6, §7).** The manifest's new `live` member carries the frame's room-clock time,
  its RTP time and, on a group's first object, the grant reference of live-rooms §4.5. RTP
  time maps to the room clock from the track's first frame, re-anchored past `max_skew_ms`,
  so every frame's time lies within that bound of its arrival at the gateway.
- **Receipts (§8).** The gateway signs each group's receipts as the first hop, after the
  group's last object, chained per track on an alias derived per session. They state what
  the gateway received, not what the publisher signed.
- **WHEP (§9).** An egress session reads the owner's catalog once its receipt verifies,
  joins each track at its latest group, and verifies every object (its receipt's group and
  Object ID, the signature under a trusted first hop, the chain, the inclusion proof, the
  envelope) before the frame reaches the player; what fails is dropped and counted. A WHEP
  player sees no receipt; a WAI-aware player subscribes over MoQ and verifies itself.
- **Mapped, not yet implemented (§12, §13).** E-RTMP v2 multitrack: `capsEx`, the stream
  name as the token, tracks by the grant's profile layers, FourCCs to capabilities, the
  reconnect request. SRT 1.5.7: a one-time session ID from an HTTP exchange in the stream
  ID's `s` key, rejection codes from the HTTP statuses, MPEG-TS stream types to
  capabilities, latency.
- **How an ingest relates to a take (§15)** and **to metering (§16)**: an ingest is the live
  tier, never a `single-take`; a device that speaks capture binding still commits its take
  beside its WHIP session. Each gateway session ends with a signed record in the shape of
  live metering's per-hop record.
- **Reference.** `joule-moq-ingest` in `joule-code/joule-code-rs/crates/`, on `str0m`
  (sans-IO WebRTC, its pure-Rust crypto backend). Its reading of rooms and grants is held to
  `live-conformance/rooms/`.

### `wai.video.vp8` and `wai.audio.opus.packet`: the media of a WebRTC ingest

- **Why.** A live ingest of WebRTC media hands WAI the codecs WebRTC endpoints send.
  VP8 is mandatory to implement in every WebRTC browser, and in every WebRTC
  non-browser that sends or receives video (RFC 7742 §5); a live track carries
  Opus one packet per object, not as an Ogg stream. SPEC §5 *Broadcast and
  streaming codecs* registers both, 2026-10-18, with no edition named.
- **`wai.video.vp8`** (RFC 6386): VP8 frames in the IVF frame container with
  FourCC `VP80`, a 32-byte file header and a 12-byte header before each frame, the
  framing `wai.video.vp9` uses. Royalty-free, as VP9 (the reference source under a
  BSD licence with a patent grant). `StandardUnclassified`: RFC 6386's
  decoder-conformance text was not read, so the registry promises no byte identity;
  VP8 shares VP9's framing, not its class.
- **`wai.audio.opus.packet`** (RFC 6716): one Opus packet (§3), its TOC byte first,
  with no container, decoded with an Opus decoder (§4) of one or two channels at any
  rate it supports. It carries no pre-skip and no output gain, which are Ogg Opus's
  (RFC 7845 §5.1); `wai.audio.opus` remains the capability for a stored stream. A
  different framing takes a string of its own (§8), suffixed as
  `wai.video.av1.lossless` is. Royalty-free; `StandardTolerance`, as `wai.audio.opus`
  (RFC 6716 §6). How consecutive objects of one live track decode is the live
  binding's rule ([live-ingest](extensions/live-ingest.md)).
- **Neither is a default.** Both are standalone, `Lossy`, sink-supplied, and outside
  the floor and the recommended set; the native sink advertises neither. Neither is a
  base of `wai.video.lcevc`. The codec glossary links its VP8 and Opus entries to
  them, and the research registry records the licence and replicate basis of each
  (`research/media-codec-registry.md`, ¹⁵). The registry export, the vocabulary and
  the table in `REGISTRY.md` are regenerated.

### Live delivery hardening: receipt chains, the receipts' binary form, room track names, proof of possession

- **A join's fill from the chain's first group.** [jwp-receipts](extensions/jwp-receipts.md)
  §3 step 3: below a join's anchor a group with a `null` parent is the chain's true start,
  `parent_not_held` and held as history until the group above it links back to it, not a
  second chain head (`chain_break`). A `null` parent outside such a gap, and a fork, are
  still `chain_break`. `wai-rs`, `jwp_verify.py` and two corpus vectors
  (`chain_join_fetch_from_group_0`, `chain_join_second_start_after_the_fill`, §5.9).
- **The receipts' binary form.** jwp-receipts §2.7 (new): group and object receipts, and a
  group's receipts together (`WJR1`), as bytes. A group is its canonical bytes (§3 step
  3), its link hash computed over them as received; an object its members in fixed widths.
  Reading yields the JSON form's members, and §3 applies to them unchanged; §2.7.4 says
  what does not read. [moq-streaming-format](extensions/moq-streaming-format.md) §4.1: a
  receipts-track object carries this form, not JSON. 20 vectors (procedure `binary`,
  §5.12), read and written byte for byte by `wai-rs` (`jwp_delivery::binary`),
  `jwp_verify.py` and `joule-moq`.
- **Room track names.** [live-rooms](extensions/live-rooms.md) §6.1 names a media track's
  receipts track (its name followed by `_receipts`) and a publisher's catalog track (the
  room and owner fields, Track Name `catalog`); §6.2 ties the catalog's
  `wai_receipts_track` to it; §10.2's `moqt` scopes admit both. The rooms corpus's scopes
  change accordingly.
- **Proof of possession.** live-rooms §10.5 (new): a grant is public, so every binding
  honours it only beside a `WRP1` proof by its principal over a fresh challenge the
  verifier chose and a context naming the endpoint (a verb and `//authority/path`). Over
  MOQT the relay's SETUP carries the challenge (`GRANT_CHALLENGE`) and the client's the
  grant and the proof (`GRANT_PROOF`); a SETUP whose grant has no proof that passes is
  closed `UNAUTHORIZED`. Over HTTP (WHIP and WHEP's Bearer token) the proof rides
  `WAI-Room-Proof`, and single-use challenges `WAI-Room-Challenge`. §10.5.4's checks
  add eight reasons to §11 (`pop_missing` to `pop_wrong_context`); the CAT mapping and
  §13 follow. The rooms corpus gains 17 proofs, MOQT and HTTP; `live_rooms_verify.py`
  checks them from the text.
- **Registry.** Wire identifiers `WJR1`, `WRP1`, the Setup Options `GRANT_CHALLENGE` and
  `GRANT_PROOF`, and the HTTP fields `WAI-Room-Proof` and `WAI-Room-Challenge` (two new
  kinds, `setup-option` and `http-header`), the last four proposed and not registered.

### `wai.live.metering`: a room session's units and joules per hop

- **The extension** ([live-metering](extensions/live-metering.md)) registers the record
  `wai.live.metering`. Every hop that carried a room session, a relay, a compositor or an
  ingest gateway, signs one **per-hop session record**, `WMH1`: per track and direction the
  objects and bytes it carried and the objects it dropped by reason (delivery timeout,
  refused at the door, duplicate, queued at the end) with those whose receipts it
  forwarded; its subscriber sessions' time by layer class; the events it delivered; its own
  energy with the acquisition class of energy-measurement, or the unmetered marker; and the
  root of a **ledger**, one entry for each grant its sessions ran under, with that grant's
  joule and funds ceilings. Media tracks are counted one by one, events tracks and every
  other track of the room together, subscribers by layer class, so a record is at most
  552 + 180 × (*M* + 2) bytes, *M* the media tracks the room published, whatever the
  audience; a record past that is refused before it is read.
- **The room-session receipt**, `WMR1`, signed by the room authority or by a key the
  origin's key log lists with the new authority `live/meter` (and `energy/<class>` for each
  class it totals): a key-log authority rather than a grant power, since a receipt states
  every principal's consumption and a grant's powers are live-rooms' closed set. It names
  the per-hop records, the takes and the composition receipts it aggregates and states
  publisher-seconds, subscriber-seconds by layer class, composition-seconds, events sent
  and delivered and recorded-seconds (the capture each device committed while capturing),
  every hop's joules with its class, one total per class and never one across classes
  (refused `classes_summed`), and each grant's consumption: its units, a share of each
  metered hop's figure by the bytes it moved, labelled with that figure's class, its funds
  under a stated tariff, and its verdicts against its ceilings, committed by a consumption
  root and every grant past a ceiling named. A verifier recomputes every number from the
  records it names; eighteen refusal reasons.
- **QoE.** Clients' delivery-telemetry observations, as CMCD version 2 Event-mode reports,
  ride the CMCD event timeline of `draft-wilaw-moq-cmcd-event-timeline-00` (`T` the
  report's `ts`, values as strings, `sf` and the custom keys not carried), or CTA-5004-B's
  `application/cmcd` batch, whose line syntax the reports already have. Only a cell of at
  least fifty sessions enters the receipt, suppressed below that; the observations never
  leave the session and the room's service.
- **C2PA.** `c2pa_emit` (with `live_metering`) writes a receipt into a real signed manifest:
  `c2pa.environmental-sustainability` with `energy_kwh` only where one class covers the
  session whole, GHG and water only when the caller supplies them with their method, and
  always the companion `science.transaction.wai.live.metering` with the totals by class.
- **The reference relay** signs the per-hop record of a room from its session summaries
  (joule-code's CHANGELOG), and the corpus holds a record it signed in its test.
- **Reference and corpus.** `wai-rs` feature `live_metering` (`src/live_metering.rs`;
  `capture_take::find_take` exposes capture-binding §9.2 to it); `live-conformance/metering/`,
  written from fixed seeds and the relay's recorded record by `wai_live_metering_vectors`, and
  checked by `tests/live_metering.rs` and by `live-conformance/live_metering_verify.py`, a
  second implementation that shares no code with `wai-rs`, verifies rooms, takes and
  composition receipts with their corpora's own second implementations, recomputes every
  unit, total and consumption entry, and requires every mutated copy of the corpus to fail.
  The registry lists the record and the magics `WMH1` and `WMR1`.

### wai-quantum 0.6.4: noisy Clifford circuits by Pauli frames, the state vector on the GPU, and the density matrix by superoperators

Additive. Every pinned digest is unchanged.

- **`DensityMatrix::evolve`** (`quantum_noisy`): the exact noisy state on up to 16 qubits
  (`MAX_EVOLVE_QUBITS`), against 12 for `DensityMatrix::run`, which stays as the referee.
  - Each gate goes on `ρ`'s rows, its conjugate on the columns. Each channel becomes one
    superoperator `Σᵢ wᵢ Kᵢ ⊗ Kᵢ*` on its qubits' row and column bits, in `run`'s order.
  - A gate and its noise fuse into one gate on at most four of the vector's qubits, applied
    in the state vector's tiled passes. A gate therefore costs one pass and no copy of `ρ`.
  - **Checked:** against `run`, every entry within 10⁻¹² on circuits with every channel and
    a Toffoli. At 14 qubits, past `run`'s size, against branching trajectories within four
    standard errors.
  - **Measured** on a 16-core desktop processor: 12 qubits in 0.53 s against 158 s for
    `run`, and 10 qubits in 39 ms against 4.1 s. 15 qubits with 132 noisy gates take 49 s.
  - **The service's `density_matrix`** runs on it, up to 16 qubits within a job's memory and
    a work cap. `auto`, with noise, still tries it first.
- **`quantum_spd::expectation_capped`**: the sparse-Pauli engine held to a number of terms.
  Whenever a rotation leaves more, the threshold doubles and the sum is truncated again
  until it fits, and everything dropped that way joins the bound. A run that never reaches
  the cap is the uncapped run, bit for bit.
  - **Checked:** on the 127-qubit kicked-Ising circuit, a cap above the peak changes no bit.
    A cap a tenth of the peak stays within its bound of the uncapped value. On a small
    register, the bound covers the exact error.
  - **The service's `pauli`** holds as many terms as the job's memory allows (`max_terms`
    can lower it). A sum that would have outgrown the job now finishes at a coarser
    threshold, reported as `threshold_reached`, with its bound.
- **The tensor-network executor blocked and threaded** (`quantum_tn`). Each pairwise
  contraction is a complex matrix product blocked for the cache, its rows shared between
  threads. Slices running in parallel keep their contractions on one thread each. Every
  output adds its terms in the same order, so the result is the same bits (a test holds it
  to the plain loop's, and the pinned digests hold).
  - **Measured:** large products went from 2.0 to about 25 billion complex multiply-adds a
    second on a 16-core desktop processor (1024 × 1024 × 1024: 0.52 s → 0.043 s).
- **Faster state-vector kernels** (`quantum_sv`). Gates on two to five qubits run at
  compile-time sizes, several groups side by side, with groups of low strides gathered.
  Every output is the same sum in the same order, so the state is the same bits.
  - Measured in cache on one core of a desktop machine: two-qubit gates 280–345 → 400–460
    million amplitudes a second, and dense three- and four-qubit gates 10–25% faster.
  - On a 16-core desktop processor, dense three- to five-qubit gates are 35–50% faster.

- **`quantum_sv_metal`** (new feature, macOS): the dense state vector in single precision
  on the platform's GPU, in one buffer the GPU and the host share. Half the bytes of the
  double-precision state, so one more qubit fits.
  - Gates are fused as `quantum_sv` fuses them, then grouped into tile passes. Each
    threadgroup copies 4096 amplitudes into threadgroup memory, applies every gate of the
    pass there between barriers, and copies them back.
  - A kernel is built per group size, so each group stays in registers. The kernels are
    compiled with the platform's safe math mode. Norms, expectations and samples are summed
    on the host in double precision.
  - **Checked:**
    - Against the double-precision state vector on random circuits of 6 to 21 qubits at
      every fusion width: amplitudes within 2·10⁻⁶, expectations within 10⁻⁵.
    - Tiled and untiled runs are the same bits, and repeated runs are the same bits.
  - **Measured** on a desktop machine's GPU, with the machine heavily loaded by other work:
    a 30-qubit, depth-20 brickwork of 910 gates runs in about 4.2 s at fusion width 3 with
    tiles, against about 8 s gate by gate.
  - **The service** offers it as `statevector_single`, accuracy class `single`, where a GPU
    is present. `auto` tries it after the double-precision `statevector`, so it takes the
    circuits that one cannot hold: up to 33 qubits within a job's memory.
    - **Measured** through the public service, on a host under heavy load:
      - a 33-qubit job (a 64 GiB state, 97 gates, chosen by `auto`) in 70 s;
      - 31 qubits in 23.5 s;
      - 26 qubits in 2.1 s, against 10.1 s for the double-precision method. The expectation
        values agree to 10⁻⁸.

- **`quantum_tableau::frame_counts` and `frame_expectations`.** Clifford circuits under Pauli
  noise (`FrameStep`: depolarizing on one or two qubits, dephasing; readout flips), on up
  to 32768 qubits.
  - One noiseless outcome from the tableau is the reference. Each shot carries a Pauli
    frame, the difference its noise makes, 64 shots to a machine word.
  - Frames start with random Z bits. They leave `|0…0⟩` unchanged, and carried through the
    circuit they spread each shot uniformly over the noiseless outcomes.
  - Errors are placed by geometric gaps, with the crate's own logarithm.
  - An expectation value is the noiseless value times the frames' mean sign against the
    string, with its standard error.
  - Both readouts are the same bits at any thread count.
- **Checked:** on random Clifford circuits of 3 to 5 qubits with every channel, the counts
  (200,000 shots, with readout flips) and the expectation values agree with the exact
  density matrix within five standard errors. Two mutations are caught: dropping one qubit
  of the two-qubit channel, and dropping the random Z bits.
- **Measured** on a 16-core desktop processor: 100,000 full-width shots take 44 ms at 1,000
  qubits with 10,000 noisy gates, and 1.6 s at 10,000 qubits with 100,000.
- **The service's `stabilizer` method takes Pauli noise.** Each QASM gate gets its noise on
  the qubits it touched, as the other noisy methods apply it. Expectations are estimated
  from at least 10,000 shots unless `trajectories` says otherwise. Amplitude damping is not
  a Pauli channel and is refused. With noise, `auto` tries `density_matrix` (12 qubits,
  exact), then `stabilizer`, then `trajectories`.
- **The OpenQASM writer keeps every gate** (`quantum_toolchain::to_qasm`). A gate with
  three or more controls, such as Grover's oracle and diffusion past three qubits, was
  written as a comment, so the program written out was not the circuit. Every controlled
  gate without a standard name is now written in OpenQASM 3's modifier form,
  `ctrl(n) @ g`, and `from_qasm` reads `ctrl`, `ctrl(n)`, `negctrl` and `negctrl(n)` back.
  - **Checked:** each base gate under zero to five controls is read back op for op, and
    Grover's search on five and seven qubits keeps its state.
- **The service** (`wai-quantum-service`, not published):
  - It reads the control modifiers on `x y z h s sdg t tdg p` and on `cx cy cz ch cp ccx
    ccz`. A gate with a standard name is read as that gate, so a Clifford program keeps
    the stabilizer method. A phase on m qubits is written exactly with 2^m − 1 `cx` and
    as many `p`, its angles dyadic when the phase is. **Checked:** on a 6-qubit state,
    every amplitude against the gate applied directly where the controls hold, within
    10⁻¹², each test failing when the expansion's signs or its releases are broken.
  - Receipts sign the RFC 8785 canonical text of the request and the result, so a result
    verifies after any JSON reader has carried it, a browser's included. Its numbers are
    checked against a JavaScript engine on RFC 8785's appendix B and on 2 million random
    doubles. Receipts made before it still verify.
  - It reads numbers correctly rounded (serde_json's `float_roundtrip`). Before, a result
    could change in its last bit on its way from the job's process to the reply.
  - A job carries its request as submitted, so whoever holds its id can check its receipt.

### `wai.image.jpegai`'s hyper decoders, context stages and synthesis through the integer operator set

- **§13.6.2 is normative.** [integer-operators](extensions/integer-operators.md) §13.6.2
  writes the topology of the capability's other integer networks as four `WHG1` graphs,
  lowered by `jpegai-int/1` from the bundle's eight weight files in §13.0's canonical
  form: `rec_y` and `rec_uv` (each plane's hyper decoder, and for luma its four context
  stages, from the hyper-latent and the dequantised residual to the reconstructed
  latent) and `syn_y` and `syn_uv` (the synthesis). §1's claim widens to what these
  networks give for their inputs. The independent checker (`tools/wai_ops_check.py`),
  written from the text alone, reproduces every lowered graph and certificate byte for
  byte.
- **Ranges.** Every activation an operator can grow, and every input but the
  hyper-latent, is checked within i32 (`A32`), so every sum is bounded by the weights.
  On the test codestream inputs stay within 21 bits, activations within 24, and the
  largest sum is 37 bits against certified bounds of 42 to 60: `acc64` holds for every
  sum, and no wider accumulator is needed (§13.6.2, *Accumulators, measured*).
- **Profiles.** The `ops` member is `{ wai-ops/1, acc64, full, jpegai-int/1,
  role_profiles: { hs_y: acc32, hs_uv: acc32 } }`: the set meets `acc64`, and its scale
  path still `acc32` (§11's new `role_profiles`; both certifiers take
  `--role-profile`). The certificate of `jpegai/v1` (`3e243dc1…`) holds the six graphs.
- **The switch.** `jpegai_decode::{decode, decode_with_stages}` run the hyper decoders,
  context stages and synthesis through the operator set (`jpegai_ops::Networks`,
  `jpegai_decode::load_networks`); `decode_loaded` takes the scale path and the networks
  loaded once, with the eICCI net. Every committed and pinned output is unchanged: the
  `Synthesis` stage hash and every other stage hash of `tests/jpegai_cross_platform.rs`,
  and the decoded RGB SHA-256 that `wai-jpegai-e2e.yml` pins. The direct networks run
  only in `jpegai_decode::legacy`, test-only (`legacy_int`). The neural sink lowers and
  loads the networks from the members' bytes through the store of loaded sets.
- **No loader panics.** `Weights::from_json`, `IntNet::from_json` and
  `IcciNet::from_json` return a `Result` (`jpegai_synthesis::BadMember`); a mistyped
  member is refused `model_mismatch` where it panicked, and the sink reports that
  refusal rather than decoding through the float path.
- **Planned breaks** (integer-operators §15.3 item 4, SPEC §8): a weight file that
  breaks §13.6.2's member rules or shapes, or networks whose certificate does not meet
  `acc64` (`model_mismatch`); a dequantised residual, a latent or an activation of the
  networks past i32 as a Q16 value (`field_range`), where the decoder computed on. No
  committed bundle or codestream reaches either.
- **CI.** `tests/jpegai_nets_ops.rs` runs in both fast-kernel steps of
  `wai-conformance.yml` on the committed synthetic networks, and in `wai-jpegai-e2e.yml`
  on `jpegai/v1`, each pinned case required by name; the differential gains the
  networks' half, per push and nightly.

### wai-quantum 0.6.3: the wide stabilizer tableau, matrix product states at any angle, and fewer passes for the state vector, noisy shots and sparse Pauli sums

Every pinned digest is unchanged, and the state vector and the sparse-Pauli engine give the
same bits as before.

- **`quantum_tableau`** (new feature, in `full`): Clifford circuits on up to 32768 qubits.
  The Aaronson–Gottesman tableau is packed qubit by qubit, with 32-bit qubit indices, so a
  gate costs a few word operations per 64 rows.
  - **Outcomes:** `Tableau::support` computes the affine subspace the computational-basis
    outcomes are uniform over, by Gaussian elimination on the stabilizers with their signs
    tracked. A `Sampler` then draws each shot as a random combination of its basis,
    projected onto the measured qubits.
  - **Expectation values:** `Tableau::expectation` gives the exact ±1/0 of a Pauli string,
    from the destabilizers.
- **`quantum_mpsim`** (new feature, in `full`): matrix-product-state simulation of circuits
  with any angles, on up to 4096 qubits.
  - The chain is kept in mixed-canonical form, the center moved by QR. A gate on distant
    qubits is routed by swaps.
  - A two-qubit gate is a two-site update. It keeps the leading eigenvectors `U` of `M·M†`
    and makes `U†·M` the new center, so no singular value is divided by. A matrix taller
    than wide is first factored `M = Q·R`. The update is truncated to the bond dimension and
    a cutoff.
  - The eigensolver is new in `linalg`: Householder reflections to a real tridiagonal form,
    then QL with implicit shifts, in a fixed order with only `+ − × ÷ √`. Shares below 10⁻¹²
    of an update's weight are its rounding and are always dropped. An update that loses
    under 10⁻¹³ of its weight is not counted as a truncation.
  - Every truncation is reported: the bonds used, the summed discarded weight, and the
    fidelity estimate `Π(1 − εᵢ)`.
  - Readouts: amplitudes, Pauli expectation values by transfer matrices, and shots by
    sequential conditional sampling.
  - **Measured:**
    - With nothing truncated, every amplitude and expectation agrees with the dense state
      vector to 1e-10, and sampled frequencies match the probabilities. Under truncation,
      the estimate tracks the true fidelity.
    - On a 16-core desktop processor, on nearest-neighbour brickworks, against the crate's
      one-sided Jacobi split: 500 qubits at depth 10 take 0.75 s against 3.1 s, and depth
      20 at bond 64 takes 8.1 s against 35.9 s.
    - At depth 20 the estimate reports what truncation cost: 0.0426 at bond 64 and 0.4712
      at bond 128, the same under both splits.
- **Measured** on a random Clifford circuit of 10n gates at 10,000 qubits: the gates in
  49 ms, the outcome subspace in 353 ms, 1000 full-width shots in 154 ms, one expectation
  value in 27 ms. Over 40 random circuits on up to 10 qubits, the outcome set and its
  uniform probabilities match the dense state vector's, as do 800 Pauli expectation values.
- **Tiled state-vector passes** (`quantum_sv`). `StateVector::run` now groups consecutive
  gates into passes, as a `Schedule` sets (`run_scheduled`; `Schedule::UNTILED` gives a pass
  per gate). Each tile holds the amplitudes that agree on every bit outside the pass's local
  bits. A tile is copied into a buffer the size of a core's cache, every gate of the pass is
  applied to it there, and it is copied back.
  - A pass's dense gates act on its local bits. Diagonal gates may act on any bit.
  - Gates keep their order, and each amplitude goes through the same expressions as before,
    so the schedule never changes a bit. A test holds tiled and untiled states equal bit for
    bit at several tile and run sizes, at one thread and many.
  - A gate of three or more qubits on the top qubits used to fall back to one thread; it now
    runs in tiles. Gates on two or more qubits whose lowest stride allows it update eight
    groups side by side.
  - **Measured** on a 16-core desktop processor, 26 qubits, double precision:
    - a depth-20 brickwork of random rotations and fSim gates: 5.2 s → 2.3 s at fusion width 2;
    - a QFT: 2.9 s → 1.6 s at width 3;
    - a QAOA circuit: 2.9 s → 0.95 s.
- **Branching trajectories** (`quantum_noisy::branching_trajectories`). The shots of a noisy
  circuit have the law of independent trajectories, but shots that took the same errors
  share one state.
  - At each noise event a group splits as independent shots would. The number that err is
    binomial (drawn by geometric gaps, so the cost is the number of errors, not of shots),
    and each erring shot chooses its Pauli uniformly. A damping event's jumps are binomial
    in the group's own jump probability. The error-free prefix and the error-free majority
    therefore run once.
  - The largest branch carries on in place. Other branches run first in copies while fewer
    than a stated number of states are alive; past that, a branch waits and is replayed.
    Estimates, counts and histories depend on neither the thread count nor the state limit.
  - **Checked:** expectations land within four standard errors of the exact density matrix
    for three seeds, and counts with readout flips follow the exact distribution.
  - **Measured**, 18 qubits and 300 shots of a brickwork circuit, against independent
    trajectories: 28 times faster at two-qubit error 10⁻³, 5.6 times at 5·10⁻³, and 1.9
    times at 2·10⁻² with damping.
- **The sparse-Pauli engine on a worker pool** (`quantum_spd`). A large sum is updated by
  workers started once per run, each taking pieces of the table as it frees up.
  - A run of consecutive Clifford gates is applied in one pass, each term through the gates
    in order.
  - A rotation's new terms have distinct strings, so each existing coefficient receives at
    most one addition. The rest are appended in table order, and dropped magnitudes are
    summed in table order.
  - The thread count therefore changes the time and never a bit: a test compares one, two,
    three and eight threads bit for bit. `expectation_threaded` takes an explicit count.
  - **Measured:** the 127-qubit kicked-Ising run (20 steps, threshold 10⁻⁴, 2.2 million terms
    at its peak) takes 9.4 s, against 49 s on one thread, with the same value.

- **The tensor-engine member's kernel compiles for a device.**
  - **The kernel compiled for the simulator only.** The helper
    (`quantum_fabric/tensor_engine_helper.py`) defined its kernel inside `main()` and
    imported the kernel toolchain there. The simulator runs the kernel as Python, so it found
    the names. Compiling for a device traces the kernel's source and looks its names up in
    the kernel's module, so it found none and refused the first request (`failed to resolve
    name`). On a device every test of the member failed at that request. The kernel now sits
    at module level. `main()` still imports the toolchain after the target is chosen, but
    binds it as module globals.
  - **The compiler is found beside the interpreter.** Compiling for a device runs the
    toolchain's compiler from `PATH`, and the interpreter the member is given need not be on
    it. The helper now puts that interpreter's directory first on `PATH` for a device target.
  - **Checked** off the device with the toolchain's own compiler: the kernel now compiles
    for the device target. Execution there needs the device's runtime. The simulator path is
    unchanged.

### wai-quantum-service 0.1.0: the quantum stack as a public job service

- **What it is** (`wai/wai-quantum-service`): an HTTP job service over wai-quantum. A job is
  an OpenQASM 3 or 2 program and the outputs wanted: measurement counts, Pauli expectation
  values, or amplitudes of named bit strings, optionally under a noise model. The service
  answers with the method that ran, its accuracy, and a receipt. No account and no key.
  The public instance is at `https://wai.transaction.science/quantum-jobs/v1`.
- **The OpenQASM front end** (`src/qasm.rs`) reads:
  - registers, and broadcast over them;
  - user `gate` definitions, inlined with their parameters;
  - angle expressions;
  - the standard gate libraries;
  - every terminal-measurement form.

  Mid-circuit measurement, classical control and gate modifiers are refused, with their
  line. It lowers one circuit three ways, each exact up to a global phase: to matrix gates
  (the dense state vector, tensor networks, noise), to Pauli rotations (sparse Pauli
  dynamics) and to the dyadic gate set (the bit-exact state vector, the stabilizer tableau).
  The tests hold each lowering to the others on every gate.
- **Seven methods, and `auto`:**
  - `exact` (bit-exact, dyadic angles, 26 qubits);
  - `statevector` (exact, any angles, up to the memory a job has: 32 qubits);
  - `stabilizer` (Clifford, 32768 qubits, on the wide tableau below: counts and exact
    expectation values);
  - `pauli` (bounded, 1024 qubits, each value with its truncation bound);
  - `mps` (any angles, 4096 qubits, at bond dimension up to 256: exact when nothing is
    truncated, otherwise heuristic with its fidelity estimate and a warning below 0.9);
  - `tensor_network` (exact amplitudes, with the contraction's cost);
  - `density_matrix` (exact noise, 12 qubits);
  - `trajectories` (noise as an estimate with standard errors).

  `statevector` runs in tiled passes. `trajectories` runs every shot as a branching
  trajectory and reports how many distinct error `histories` ran. `pauli` reports a
  `ladder`: each value and bound at up to three thresholds a decade above the one asked
  for (`"ladder": false` turns it off). `auto` picks the first that takes the job, and the plan says why, and why not the others.
- **A public service on a shared host:**
  - every job is planned before it is queued, and one whose memory estimate is over the
    per-job cap is refused at once;
  - jobs run inside a total memory budget, and a large one waits until the host's free memory
    covers it and a reserve;
  - each job runs in its own child process, at a lower priority than the host's other work,
    watched for its resident memory and its time, and stopped past either;
  - clients get per-client active-job and submission limits, against a bounded queue.

  The limits are stated at `GET /v1`.
- **Receipts:** every finished result is sealed with the service's Ed25519 key over the
  BLAKE3 digest of the request and the result; `POST /v1/verify` checks one.
- **Measured:** 29 tests on Linux and macOS, and `smoke.sh` against local instances and the
  public one. That covers a 40-qubit GHZ by the stabilizer, receipts that verify while a
  changed result does not, a 100-qubit circuit's expectation values by sparse Pauli dynamics
  with their bounds and ladder, 2000 noisy shots of a 14-qubit circuit by branching
  trajectories, refusals with reasons, the per-client limit, and a job stopped when it
  outgrew its memory. CI: `wai-quantum-service` (Linux and the macOS runner).

### `WIV1` and `WIR1` through the integer operator set's stage operators

- **`ops` members with no graph.** `wai.video.int_motion` and
  `wai.{image,video,audio,splat}.int_refine` decode through the stage operators of
  [integer-operators](extensions/integer-operators.md) §13.5 alone, and register
  `{ wai-ops/1, stage: rans64 t4097/lenient/canonical }`: with no parameter set, the
  member has no profile, subset or lowering (§11; the registry schema now requires
  `set`, those three together, and a lowering or a stage).
- **The stage operators the drivers need.** `wai-ops`' E1 decodes a symbol against a
  table in whatever width a payload's parser keeps it (`Rans::symbol_in`, and
  `table_ok` reads its entries likewise), and E10, E11 and E12 gain per-element forms
  (`mc_predict_at`, `up2_at`, `rate_at`) that read a frame, a picture or source samples
  in the driver's layout, so a decode forms no copy.
- **The switch.** `int_video::{decode_typed, decode_pframes_typed}` decode `WIV1`
  through E1, E10 and E9 (`int_video::ops`), and `staged::wir1::apply_unit` and
  `Wir1::check_streams` decode `WIR1` through E1, E11 or E12, and E9
  (`staged::wir1::ops`). Every committed output is unchanged: the `int_video` clips
  (the half-pel real footage among them) and their SHA-256 sidecars, the payload
  corpus, every stage of the staged corpus and the byte-equality goldens. The direct
  decodes are `int_video::legacy` and `staged::wir1::legacy`, test-only (`legacy_int`).
  No input decodes or refuses differently ([integer-operators](extensions/integer-operators.md)
  §15.3 item 5).
- **The 32-bit form, behind a feature.** `wai-ops`' E9 32-bit form
  (`stage::residual_step_i32`, `i32_form_fits`, and `residual_step_by_i32` for `i64`
  operands) is behind the new `i32_forms` feature, the arithmetic a 32-bit fabric runs;
  `wai-rs`' `i32_forms` decodes `WIV1` and `WIR1` through it
  (`int_video::ops::decode_typed_i32`, `staged::wir1::ops::apply_unit_i32`), held to the
  exact decode by the parity tests, the differential and the byte-equality goldens.
- **Text.** [integer-operators](extensions/integer-operators.md) §5E writes E10, E11 and
  E12 out, §13.5 the `WIV1` and `WIR1` drivers, and the status note names the stage-only
  capabilities; integer-payloads §3 and §12 and staged-delivery A.4 cite the stage
  operators; SPEC §8 lists no planned break for them.

### `wai.image.jpegai`'s scale path and entropy decodes through the integer operator set

- **The residual-variance bands start where the reference decoder's tables change.**
  The exact decoder put a block whose `likely` was exactly 1003, 1352 or 1937 in the
  next band, for its σ addition and for its dequantisation multiplier. The reference
  decoder's own tables (4352 entries per branch, indexed by `likely`) change value at
  1004, 1353 and 1938, and the reference sink's float decode path, which reads those
  tables, already followed them. The edges are now 1004, 1353 and 1938, in the decoder
  and in [integer-operators](extensions/integer-operators.md) §13.6.1 step 3 (§15.3
  item 4). `tests/jpegai_mans_golden.rs` holds both branches' additions and multipliers
  to the reference's tables at every index. No committed codestream has a block on an
  edge, so no output changes.
- **The lowering, from the bundle the decoder holds.** `wai-ops` gains
  `lower::parse_jpegai`, `lower::jpegai_net` and `lower::jpegai_net_loaded`
  (`jpegai-int/1`, [integer-operators](extensions/integer-operators.md) §13.6.1, of a
  bundle already parsed; the last loads both scale graphs without writing and parsing
  them again), `lower::jpegai::scale` (one plane's graph) and
  `stage::mans::decode_sgm_tables` (E8 with its tables built once). The committed graphs
  and certificates are unchanged.
- **The switch.** `jpegai_decode::{decode, decode_with_stages}` run the scale path and
  the two entropy decodes through the operator set (`jpegai_ops`): the bundle's scale
  graphs loaded under `acc32` with its tables (`jpegai_decode::load`, a
  `jpegai_ops::ScalePath`), then each plane's hyper-latent, scale, mask, `likely`, RVS
  addition, `σ_idx` and latent. Every committed and pinned output is unchanged: the
  decoded RGB of `jpegai/v1`, every stage hash of `tests/jpegai_cross_platform.rs`, and
  the reference decoder's symbols and scales. Each now returns a `Result`
  (`jpegai_decode::Refusal`); `decode_loaded` decodes with a scale path loaded once. The
  direct decode is `jpegai_decode::legacy` and `jpegai_ops::legacy`, test-only
  (`legacy_int`). `NormTables`' bounds and CDF rows are `i64`, so a load check sees the
  value the file gives. The float decode path decodes its scale path and entropy through
  the same driver, from its own bundle's members.
- **`ops` member.** `wai.image.jpegai` registers `{ wai-ops/1, acc32, full,
  jpegai-int/1 }` (SPEC §5, §7).
- **No silent fallback.** The neural sink lowers and loads the scale path from the
  members' bytes through the store of loaded sets (keyed by the listing of `norm.json`
  and `weights/hs_{y,uv}.json`). A refusal of the exact decode is the decode's result
  (`model_mismatch`, `stream_exhausted`, or `set_incomplete` for a member the set
  lacks); only a sink without the bundle, or a thread or target that cannot run the
  float stages, decodes through the float path ([determinism-tiers](extensions/determinism-tiers.md) §5).
- **The browser crate.** `decodeJpegAiMans` decodes through E8 and throws on a refusal;
  `decodeJpegAiFull` reports the decode's refusal.
- **One implementation.** The float path's own ports of the mANS decoder and of the
  hyper-scale network (`neural::jpegai_mans`, `neural::jpegai_hyper`) are deleted: that
  path runs the shared driver. The exact decoder's direct pieces (the crate-root mANS
  port, `jpegai_synthesis::decode_hyper_scale`, and the pipeline's `skip_mask`,
  `rvs_additive`, `rvs_likely` and `sigma_index`) are test-only (`legacy_int`), the
  reference the differential (`tests/int_ops_differential.rs`) and the refusal-parity
  tests hold the operator set to.
- **Planned breaks** (integer-operators §15.3 item 4, SPEC §8): a bundle that fails a
  load check of §13.6.1 (a bound above 255 or a shift outside `0..=127`, which the loader
  wrapped, among them) or whose scale graphs do not meet `acc32` (a sum past `i32`, a
  shift past 63) is refused (`model_mismatch`); a substream that ends before the decode
  it carries is refused (`stream_exhausted`) where the decoder panicked. No committed
  bundle or codestream is affected.

### `wai.text.timed`: timed text and live captions

- **The extension** ([timed-text](extensions/timed-text.md)) registers the record
  `wai.text.timed`. A cue is a span of media time and its text in a closed set of styling,
  italic runs, line breaks and six anchors, which WebVTT and the IMSC1 Text Profile both
  express; text outside the subset both hold exactly (a tab, two spaces in a row, a space at
  a line's edge) is refused rather than changed.
- **Live captions.** A caption is a `caption` event of `wai.live.events`, unchanged. A
  machine caption is revised in place: each revision is a new `caption` event, followed by
  a `meta` event carrying the `WTX1` annotation, which supersedes the revision before it by
  digest, says whether the text is final and how much is stable, and carries the anchor and
  italics. A verifier checks that revisions form one line per cue, keep their language,
  anchor and stable prefix, and supersede only a partial of their own principal; it states
  what a sink shows at an instant and the cues a stream resolves to. A publisher repeats
  the open partial and each caption's annotation at a group's start, so a subscriber that
  joins mid-stream shows what one from the start shows. live-events needs no change.
- **Mappings.** WebVTT documents and segments with `X-TIMESTAMP-MAP`, and a reader of the
  subset; IMSC1 documents, refused when a sound check of the hypothetical render model does
  not pass; the WebVTT and IMSC1 (cue mode) MoQ caption timelines, whose first object of
  each group repeats every cue still showing, and their catalog entries; CTA-608 on field 1,
  CC1, pop-on and roll-up with windows of 2, 3 and 4 rows, odd parity, doubled codes, the
  special and extended characters with their fallbacks, and the slot clock's rate as a
  parameter; CTA-708 is specified only.
- **The disclosure label.** Origin and model digest travel through every target: the HLS
  `public.machine-generated` characteristic (`draft-pantos-hls-rfc8216bis-22` §4.4.6.1),
  C2PA 2.4 `c2pa.ai-disclosure` and the IPTC digital source type, a WebVTT class span, IMSC1
  attributes, timeline members, a catalog field, and a sidecar published by
  `EXT-X-SESSION-DATA` for targets that cannot carry it, such as 608.
- **Player obligations** for captions under the viewer's control, persistent and
  discoverable, each citing EN 301 549 V4.1.1 or WCAG 2.2.
- **Reference and corpus.** `wai-rs` feature `timed_text` (`src/timed_text.rs`, with
  `cc608.rs`); `live-conformance/captions/`, written from fixed seeds by
  `wai_timed_text_vectors` with every expected `.vtt`, `.ttml` and `.608.hex`, and checked
  by `tests/timed_text.rs` and by `live-conformance/timed_text_verify.py`, a second
  implementation that shares no code with `wai-rs`, parses every IMSC1 document with an XML
  parser, and requires every mutated copy of the corpus to fail. The registry lists the
  record, the magic `WTX1`, its catalog fields and its session data id.
- **Not checked against the standard.** CTA-608-E and CTA-708-E could not be read for this
  revision. The 608 tables are as 47 CFR 79.101 reproduces them; the extended characters are
  not there, and the extension marks its table of them for checking.

### `wai.live.recording`: a room session as one package, and clips that prove what they are

- **The extension** ([live-recording](extensions/live-recording.md)) registers the record
  `wai.live.recording`. A finished room session is one WAI package: a directory of files and
  its `LISTING` (SPEC §3.1), as a packaging package is, whose components are the room (its
  descriptor, grants and key-log head), its events (each principal's chain whole, the form
  live-events §8.6 gives a recording), one component for each participant's take (its `WTM1`
  manifest, the chunks the store holds, its receipts and consents) and, optionally, the
  composite by its composition receipts. Each component is named as SPEC §7.1 names one, by
  an `id` and a `sha256`, here the SHA-256 of the listing of its files.
- **The recording manifest** (`WRM1`) is signed by the room authority, or by a principal that
  held the record power at every instant of the session span (its grant in effect from the
  span's start, with no leave, exchange or expiry taking the power away before its end; a
  refusal states the first instant without it), and names every component with the span on
  the room clock and the region the package is kept in, held to the room's region pin. Every
  take the package holds lies within the span; the package holds the room's chains whole. A
  recording verifier replays the room and each take and gives `complete` (every take the
  room's events open is held, each one `complete` with its commitments unbroken from open to
  close, and no event missing) and `publishable` for each purpose.
- **Clips.** A clip record (`WCL1`) covers a span of one take, by chunks or by a byte range
  of the master (the boundary chunks carried whole, with trims), or one window of the room
  clock across several takes. Each part carries its chunks' descriptors and one range proof
  against a commitment the device made while it captured, the closing one or, while the take
  runs, the one that covers the span. A clip verifier needs the clip's bytes, the record, the
  take's commitments up to that one, the device's key and the room descriptor, never the take,
  and checks `single-take` over the span. An aligned clip states each part's window on its
  device's clock and its bound: the jitter to the room clock, twice it between two takes.
- **Range proofs.** The output tree gains `range_proof` and `verify_range`: RFC 6962's audit
  path taken over a run of leaves, whose one-leaf case is the audit path node for node and
  whose check agrees with RFC 9162 §2.1.3.2's, tested over every run of trees of up to 33
  leaves. The take tree reuses it; no new tree.
- **Mappings.** The HLS and LL-CMAF presentation of a CMAF master (one Media Playlist per
  take, its header by `EXT-X-MAP` byte range, parts proved with their chunk) and the MoQ
  addressing of a recorded take's chunks for a FETCH (a group per chunk, a `LOCATION_FILTER`
  per clip part) are specified; the reference states each clip part's FETCH address and
  writes no playlist.
- **Reference and corpus.** `wai-rs` feature `live_record` (`src/live_record.rs`);
  `live-conformance/recordings/`, written from fixed seeds by `wai_live_record_vectors`, whose
  performer's take is capture binding's `lesson` take byte for byte, checked by
  `tests/live_record.rs` and by `live-conformance/live_record_verify.py`, a second
  implementation that shares no code with `wai-rs`, checks range proofs by the extension's
  recursion and by RFC 9162's own algorithm, and requires every mutated copy of the corpus to
  fail. The registry lists the record and its magics (`WRM1`, `WCL1`). live-events §8.6,
  live-rooms §3.7 and capture-binding §1 point at the extension.

### `wai.live.compose`: a room's composition as an integer function

- **The extension** ([live-compose](extensions/live-compose.md)) registers the record
  `wai.live.compose`. A `WCD1` composition descriptor, signed by the room authority or by a
  principal whose grant has the compose power, fixes a canvas (each edge even, 16 to 1920
  pixels, at most 1920 × 1080 pixels, 1 to 60 frames a second), an sRGB background, the
  highest layer composed, a layout, the grant attributes it reads, and its limits on tiles
  and voices.
- **The participant set** at a room-clock instant is the room's lifecycle replay up to it:
  every present publisher but the compositor, its video source (own video, own screen, or
  its live place in a slot) and audio source, each with the dimensions of the composed
  layer, the attributes of its grant in effect, and its join order. A gap in a room key's
  chain is `lifecycle_incomplete`; a gap elsewhere changes nothing. The set has a digest.
- **The layout** (§5) is integer arithmetic throughout, with every rounding stated: a grid
  chosen by the largest tile of the reference aspect among every column count, its rows
  centred; tile aspects auto, 16:9, 1:1 and 3:4; a featured participant above the others;
  fill, cover and contain, by default cover for grid tiles and contain for the featured
  one; stopped video shown or omitted; order by attribute, then join order, then key; and
  picture in picture, the PiP participant's window over the grid in any corner, of a set or
  derived size, static or dynamic (alone, it is laid out as though it had no PiP
  attribute). The layout is computed again at every lifecycle event and grant exchange.
- **The mix** (§6): per-voice gain and pan as integers in units of 1/4096, over signed
  16-bit PCM, summed exactly in 64 bits and saturated once, so it is byte-exact over the
  decoded PCM; a decode and a lossy re-encode are not, and the receipt binds objects.
- **Where it runs** (§7): on each sink by default, with no server encode; a server
  compositor must be the descriptor's compositor, hold the compose power and sign a `WCR1`
  receipt for each object of a composed track, naming the descriptor, the set's digest at
  `t`, the output object and the input objects of exactly the tracks the layout reads.
- **Live rooms.** §1, §3.3 and §4.4 point at the extension for the compose power and the
  attributes `featured`, `order` and `pip` (and `gain` and `pan`, which it adds).
- **Reference and corpus.** `wai-rs` feature `live_compose` (`src/live_compose.rs`; the
  layout and the mixer in `src/live_compose/layout.rs` use nothing beyond `core`, `Vec` and
  `String`); `live-conformance/compose/`, written from fixed seeds by
  `wai_live_compose_vectors`, and checked by `tests/live_compose.rs` and by
  `live-conformance/live_compose_verify.py`, a second implementation that shares no code
  with `wai-rs`, steps through the room's lifecycle replay itself and requires every
  mutated copy of the corpus to fail. The registry lists the record and its magics (`WCD1`,
  `WCR1`).

### `wai.capture.take`: a device's master recording bound to a live room

- **The extension** ([capture-binding](extensions/capture-binding.md)) registers the
  record `wai.capture.take`, for the case of a lesson or an audition where the live call is
  low quality and the master is made on the performer's own device and uploaded after the
  session. A take answers a `record_start` addressed to the device; its id is derived from
  the room, the device and that event. The device's grant binds it to the principal it
  captures for (the grant attribute `subject`), and may mark that principal a minor
  (`minor`).
- **Commitments.** While it captures, the device emits a `take_commit` event every commit
  interval of the room's recording policy (live-events §4.12's payload, defined here): the
  take id, the chunk range it adds, the root over every chunk so far, its offset to the room
  clock, and an RFC 9162 consistency proof that the root extends the last one. The take
  tree is staged-measurement Appendix C's (RFC 6962's Merkle tree hash, its node) over
  leaves that bind each chunk's index, capture range, length and hash; the output tree
  exposes that hash and gains consistency proofs for it.
- **Upload and manifest.** The master is uploaded resumably by chunk hash; the store, a key
  the room's key log lists with the new authority `capture/store`, signs `WUR1` receipts of
  the upload's state machine (open, progress, resume, sealed) within the room's region pin.
  The device signs a `WTM1` take manifest listing the chunks, their root, every
  commitment's digest and each subject's `WCN1` consent (purposes live, record, distribute
  and train; a guardian's signature for a minor; revocable through the key log).
- **Verdicts.** A take verifier replays the room (live-rooms §9) and gives `single-take`,
  `aligned`, `complete` and `publishable` for each purpose a verdict, each failure a named
  reason (§12): a commitment not delivered is `delivery_gap`; a part not committed at its
  index is altered, reordered or extra; a missing part fails `complete`, not `single-take`.
  A room-clock instant maps into the master, so a marker set on the live call lands in it.
- **Live rooms.** The recording policy gains `commit_interval_ms` and `jitter_ms`
  (live-rooms §3.7; 2 000 and 500 by default), and the rooms corpus is regenerated.
- **C2PA live video** (C2PA Technical Specification 2.4 §19): the mapping of a take and its
  CMAF output onto per-segment manifests and Verifiable Segment Info, session keys, AI
  disclosure and sustainability is specified; the crate has no CBOR or COSE encoder, so it
  implements only the stream id.
- **Reference and corpus.** `wai-rs` feature `capture_take` (`src/capture_take.rs`);
  `live-conformance/takes/`, written from fixed seeds by `wai_capture_take_vectors`, and
  checked by `tests/capture_take.rs` and by `live-conformance/capture_take_verify.py`, a
  second implementation that shares no code with `wai-rs` and requires every mutated copy
  of the corpus to fail. The registry lists the record and its magics (`WTM1`, `WUR1`,
  `WCN1`).

### `wai.live.room`: room descriptors, room grants and the room verifier

- **The extension** ([live-rooms](extensions/live-rooms.md)) registers the record
  `wai.live.room`. A room is fixed by a signed descriptor, `WRD1`, whose digest is the
  room id every `WEV1` event names: its authority, its clock and `t0`, the roles a
  principal may hold (publish per media kind, subscribe up to a layer, the event kinds it
  may send, and the moderate, direct, compose and record powers), its limits and event
  ceiling, the slots several publishers may stand behind, its publish profiles, its
  review policy and its recording policy with a region pin.
- **Grants.** A principal takes part under a signed room grant, `WRG1`: its role and
  capabilities, a window of at most fourteen days, its delivery origins, a single-use
  nonce, a viewer id, the `WEP1` event policy of live-events §5 it is held to, its joule
  and funds ceilings, the profiles it publishes with, and at most 1 KB of composition
  attributes. A grant is exchanged for one chained to it at a tick of the room clock,
  recorded as the `grant_exchange` lifecycle event, and revoked through the authority's
  key log: the head's `revoked_modules` lists revoked grants and viewers, and the
  authority's key holds the new authority `live/room` (jwp-receipts §7.2, §7.5, §7.6).
- **Tracks, slots and profiles.** A track is named by the MOQT namespace tuple (room,
  principal or slot, media kind) and its layer. A slot is a virtual track: a publisher of
  higher precedence takes it over with no new session, and a standby takes it on failover
  under the slot's rule. The lifecycle kinds' `detail` is defined. The publish profiles
  `speech`, `music`, `screen`, `camera-720` and `camera-1080` fix codecs, layers and
  limits as data.
- **Room verification.** Given a descriptor, grants, a key-log head and every principal's
  chain, a verifier merges the chains by room clock, verifies each event with the
  live-events verifier under the policy of the grant in effect at its `t`, and replays the
  lifecycle into room state; each refusal has a named reason (§11).
- **Bindings**, as tables: the Common Access Token's claims (CTA-5007-B), the `moqt` claim
  of `draft-ietf-moq-c4m-01` (whose scopes the reference derives from each grant), URI
  Signing (RFC 9246) and Privacy Pass (`draft-ietf-moq-privacy-pass-auth-03`). No CAT
  encoding: the crate has no CBOR.
- **Reference and corpus.** `wai-rs` feature `live_room` (`src/live_room.rs`);
  `live-conformance/rooms/`, written from fixed seeds by `wai_live_room_vectors`, and
  checked by `tests/live_room.rs` and by `live-conformance/live_rooms_verify.py`, a second
  implementation that shares no code with `wai-rs`. The registry lists the record and its
  magics (`WRD1`, `WRG1`).

### `wai.live.events`: a signed, timeline-bound event track

- **The extension** ([live-events](extensions/live-events.md)) registers the record
  `wai.live.events`. Timed metadata, ad and chapter cues (with an SCTE 35-1
  `splice_info_section()` or an SCTE 35-2 message carried verbatim), chat, reactions,
  polls, markers, remote-control commands, captions with their machine-generation label,
  moderation decisions, take commitments and room lifecycle are each one `WEV1` record:
  signed by its principal, hash-chained per principal (`seq`, `prev`), stamped with the
  room clock and optionally anchored to one MoQ object. Its digest is SHA-256 over the
  record without the signature, and its signature is verified strictly (SPEC §7
  *Signatures*).
- **Policy and verification.** The event policy `WEP1` states the kinds a principal may
  send and the payload, text and rate bounds, every bound a default a room may change; the
  room grant of a later extension carries the same fields. A verifier refuses a record
  with a named reason (§9), and reports a delivery gap as a gap, never as corruption.
- **Polls** replay as `wai.feed.state` cells through `wai.world.sim.feed0` (`ADD_INT`):
  a tally is replay-equivalent, and a tally a poll's author states is checked.
- **Bindings**, field semantics only: an MSF event timeline over MoQ (`eventType`
  `urn:wai:event:1`, one track per principal, one-minute groups, a join set repeated at
  each group start), `EXT-X-DATERANGE` (`CLASS="science.transaction.wai.event"`,
  `X-WAI-EVENT`, with `SCTE35-OUT`/`-IN`/`-CMD` for cues), a version 1 `emsg`
  (`scheme_id_uri` `urn:wai:event:1`), a WebRTC data channel and a recording.
- **Reference and corpus.** `wai-rs` feature `live_events` (`src/live_events.rs`);
  `live-conformance/events/`, written from fixed seeds by `wai_live_events_vectors`, and
  checked by `tests/live_events.rs` and by `live-conformance/live_events_verify.py`, a
  second implementation that shares no code with `wai-rs`. The registry lists the record
  and its wire identifiers (`WEV1`, `WEP1`, `urn:wai:event:1`,
  `science.transaction.wai.event`, `X-WAI-EVENT`).

### wai-quantum 0.6.2: the HIP member's matrix-instruction kernel, and the Metal member's tensor-operation kernel

Additive. Every pinned digest is unchanged.

- **Every product by the shading language's tensor operation**
  (`quantum_fabric::metal::TENSOR_OPS_SOURCE`, `tensor_ops.metal`; `MetalMember::tensor_ops`):
  a third Metal kernel for the gate, through the language's version 4.0 tensor operation (a
  bf16 matrix multiply accumulated in f32), which the platform maps to the GPU's matrix
  hardware where it has it. The operation reads and writes device memory, so each threadgroup
  owns a slice of scratch (`TENSOR_OPS_SCRATCH_BYTES`): a 64-pair tile is one multiply of
  64 × 32 limbs by 32 × 128 weights, in four 32-column runs. Its operands are offset-binary, as
  for the HIP kernel below, so the result does not depend on how an accumulator treats a
  negative term. A device or operating system without the operation
  (`DeviceIdentity::has_tensor_ops`: GPU family 1007 or later, operating system 26 or later)
  is `MetalError::Unsupported`. Eight mutants (`TENSOR_OPS_MUTANTS`), among them nine-bit
  limbs, every run reading the first columns of the weights, and a wrong offset correction.
- **Measured** on two GPU generations, silicon: family 1009 (the macOS runner's) and family
  1010, the generation whose GPU cores carry matrix accelerators. On each: 9/9, 66020 probe
  pairs, all four pinned certificates, circuits to 22 qubits, 8 of 8 mutants killed. On family
  1010 the integer and SIMD-group matrix kernels are bit-exact too (9/9 each, 6 of 6 and 7 of 7
  mutants). With signed limbs instead, the tensor operation still agrees on all 66020 pairs on
  both generations: the offset is for the instruction below, not needed here. Where the
  platform runs the operation is its own choice; it is not observed here. CI: a step of
  `wai-quantum-metal` on family 1009; family 1010 was run by hand, the binaries built on the
  build host.

- **Every product by the bf16 matrix instruction** (`quantum_fabric::hip::WMMA_SOURCE`,
  `wmma_limbs.hip`; `HipMember::wmma`): a second HIP kernel for the same gate, for the gfx11
  architectures (`DeviceIdentity::has_wmma`), whose wavefronts of 32 multiply bf16 at
  16 × 16 × 16 and accumulate in f32. One wavefront multiplies sixteen pairs' limbs by the
  gate's limb weights in seven 16-column steps, in the instruction's own operand layout. A
  device without it is `HipError::Unsupported` (new). Eight mutants (`WMMA_MUTANTS`): nine-bit
  limbs, the result's rows in another instruction's layout, and a wrong offset correction
  among them.
- **Offset-binary limbs, because of what the instruction does.** Its f32 accumulation is not
  exact once a negative term enters a sum. Measured on gfx1151, on the instruction alone:
  `1·1 + (−1)·0` returns `1 − 2^-24`, `1 − 1` returns `−2^-24`, `128 − 64` returns
  `64 − 2^-17`; sums of non-negative terms come back exact. The CUDA kernel's signed top limb
  therefore fails the probe on this GPU, off by one in a coefficient. So each value is taken
  as `x + 2^31`, all four limbs unsigned, the instruction forms `(e + 2^31)·(a + 2^31)`
  exactly, and the kernel removes the offset in 64-bit integers:
  `e·a = (e + 2^31)·(a + 2^31) − 2^31·(e + a) − 2^62`, modulo `2^64`. The non-integer
  results the limb program gave through one compiler's HIP build for this GPU (0.6.0) are
  consistent with the same behaviour, though which instruction that build chose was not
  checked.
- **Measured** on the integrated GPU (gfx1151, silicon): 9/9, 66020 probe pairs, all four
  pinned certificates, circuits to 22 qubits, 8 of 8 mutants killed; the other kernel and the
  state vector unchanged there. CI: a step of `wai-quantum-hip`; its absent leg runs the new
  tests over no device.

### wai-quantum 0.6.1: the CUDA member's tensor-core kernel, and the Metal member

Additive. Every pinned digest is unchanged.

- **Every product on the tensor cores** (`quantum_fabric::cuda::TENSOR_SOURCE`,
  `tensor_limbs.cu`; `CudaMember::tensor`): a second CUDA kernel for the same gate, for
  devices of compute capability 8.0 or later (`DeviceIdentity::has_tensor_cores`). Each
  32-bit value is four 8-bit limbs in bf16, the lower three unsigned and the top one
  signed. One warp multiplies sixteen pairs' limbs by the gate's limb weights (sixteen
  products per pair, seven coefficients each) in fourteen `m16n8k16` steps, accumulated in
  f32. Every limb product is at most `2^16` and every coefficient sums at most four, so f32
  holds every partial sum exactly in any order. The kernel adds the coefficients, shifted
  into place, in 64-bit integers; the rest is the other kernel's arithmetic on exact
  products. The member id is the device's; the implementation id is the kernel's own
  (`cuda-tensor-limbs8/…`), and the member is probed before it is returned. A device below
  8.0 is `CudaError::Unsupported` (new). Seven mutants (`TENSOR_MUTANTS`), one of which is
  nine-bit limbs, which bf16 does not hold.
- **Measured** on the discrete GPU (compute capability 8.9, silicon): 66020 probe pairs
  agree, all four pinned certificates, circuits to 22 qubits match native, 7 of 7 mutants
  killed. The other kernel and the state vector are unchanged there (all four pins, 6 of 6
  and 4 of 4 mutants). CI: a step of `wai-quantum-cuda`; its absent leg runs the new tests
  over no device.
- **The Metal member** (`quantum_fabric_metal`, opt-in, macOS): the gate on the platform's
  GPU through its own interface, compiled at run time by the operating system's compiler.
  Two kernels, each with its own implementation id and mutants:
  - `lanes64.metal` (`metal-lanes64/…`): the CUDA kernel's exact 64-bit integer arithmetic
    in the Metal shading language, one serial dispatch per gate in one command buffer, the
    state in memory the GPU and the host share;
  - `matrix_limbs.metal` (`metal-matrix-limbs8/…`, `MetalMember::matrix`): every product
    formed by the GPU's SIMD-group matrix multiply, from 8-bit limbs in bf16 accumulated in
    f32, by the tensor-core kernel's argument. One SIMD group multiplies eight pairs' limbs
    by the gate's limb weights in fourteen 8 × 8 column blocks, and each thread joins its
    four products and forms one row. A device below GPU family 1007 is
    `MetalError::Unsupported`.

  A device is keyed by its GPU family number (`gpu-msl-<family>`), and its member id binds
  the device, its registry id, family and memory, and the operating system's version. Off
  macOS the member builds, finds nothing and says why.
- **Measured** on the macOS runner's GPU (family 1009, silicon): both kernels 9/9, each
  with 66020 probe pairs, all four pinned certificates and circuits to 22 qubits; 6 of 6
  and 7 of 7 mutants killed. Nine-bit limbs break 297 of the 484 edge pairs, so the
  multiply takes bf16 inputs. CI: `wai-quantum-metal`, on that runner, and its absent leg
  on Linux.
- **Measured by hand, outside CI:**
  - the gate as a portable tensor program, through XLA on a cloud tensor processor: the
    gate's module exact and its three mutants failing where they should, and the limb
    program (`stablehlo::limb_module`) exact on the processor's matrix unit, with the
    nine-bit set differing as it must;
  - `wai-quantum-qpu`'s `native_turns` against the service's simulator, no hardware used:
    with no noise, statistical parity at fidelity 1.0 on both gatesets' corpora (16
    circuits of 5 qubits, 1000 shots each), which checks the compilation to the service's
    native gates and the wire contract; with one of the service's device noise models,
    fidelity 0.909 (V1) and 0.927 (V2).

### `wai.world.model`'s integer renderer through the integer operator set

- **The lowering, from the models the renderer holds.** `wai-ops` gains
  `lower::world_net` and `lower::world_net_loaded`: `world/1`
  ([integer-operators](extensions/integer-operators.md) §13.5, informative) of two
  networks already parsed, the second loading `T` before it lowers `g_s`. Both write
  and refuse what `lower::world` does from the bytes; the committed sets' graphs and
  certificates are unchanged.
- **No `ops` member.** The renderer runs through the operator set, but its lowering is
  informative: the capability's registered payload is its signed action trace, and no
  payload text yet binds a trace to a renderer's set. SPEC §5's row cites §13.5, and the
  registry's test holds that `wai.world.model` has no `ops` member.
- **The direct guard, test-only.** `int_world::legacy::checked` runs the direct renderer
  with the check every other integer decoder made before each layer
  (`run_net_checked`); the tests hold the `check` ranges of `world/1` to it, layer by
  layer, at the edge of every range and through the whole renderer.
- **The switch.** `int_world::{decode, decode_deltas, decode_planes}`, and the browser's
  `WorldModel`, now render through `world/1`: `T` and `g_s` lowered and loaded under
  `acc64` once (`int_world::load`, an `int_world::World`), then planned on the latent's
  shape, then run frame by frame. Every committed frame and clip digest is unchanged, on
  every leg and in wasm32. Each now returns a `Result`; `WorldModel`'s methods throw on a
  refusal. The per-row `step`, `step_delta` and `step_planes` of a parsed model are
  replaced by `World::step`, `World::step_delta`, `World::step_action` and
  `World::render`. The direct renderer is `int_world::legacy`, test-only (`legacy_int`).
- **Planned breaks** (integer-operators §15.3 item 2, SPEC §8). Before the first frame,
  where the direct renderer rendered from the input or panicked: a set whose `afrac` is
  outside `1..=30` or unequal, with no action plane, or that breaks a `WHM1` network's
  rules (`model_mismatch`); a row without one value per action plane, and `T` or `g_s`
  that does not fit the latent (`model_mismatch`); a latent that is not `C·lat·lat`
  values (`geometry_mismatch`), of side 0 or past `2^26` values, and a move whose plane
  value leaves `i64`. A value outside a `check` range is refused (`field_range`). No
  committed set or clip is affected.
- **Memory.** With its set loaded, a clip through the operator set holds no more than the
  direct renderer held for it, and so do one step and one frame
  (`tests/int_world_memory.rs`).

### `wai.neural.int_mlicpp` through the integer operator set

- **Its `ops` member in the registry** ([integer-operators](extensions/integer-operators.md)
  §11). `registry/capabilities.json` gives `wai.neural.int_mlicpp` the member `{"set":
  "wai-ops/1", "profile": "acc64", "subset": "full", "lowering": "wmb1/1", "stage":
  {"rans64": {"table": "t2", "escape": "lenient", "end": "final_only"}, "bucket":
  "ceil"}}`. SPEC §5's row cites the payload and bundle
  ([integer-payloads](extensions/integer-payloads.md) §14) and the lowering
  (integer-operators §13.4), and SPEC §7 names the capability among those that decode
  through the operator set.
- **The decode.** `int_mlicpp::decode`, and so `int_payload::decode_int`, the neural sink
  and every sink that decodes through them, now runs the bundle's three networks through
  their lowering `wmb1/1` ([integer-operators](extensions/integer-operators.md) §13.4):
  each lowered by `whm1/1` and loaded under `acc64` with its certificate, then planned on
  the payload's shapes, before either stream. The stage steps are unchanged, with the
  `t2` table rule and the `final_only` end rule kept: a symbol decoded after a stream ran
  out is still decoded and fed to the passes after it, and the refusal comes at the end.
  The committed vector decodes to the same image, with the same bucket and symbol for
  every element; every committed refusal keeps its code. The direct decode is kept,
  test-only, behind the `legacy_int` feature. The store of loaded sets keeps the three
  loaded networks under the SHA-256 of `model.wmb`.
- **Planned breaks**
  ([integer-operators](extensions/integer-operators.md) §15.3 item 3, SPEC §8). A bundle
  is refused before either stream (`model_mismatch`) for the rules of the other `WHM1`
  decoders: a bias of magnitude `2^63 − 1` or more, a `check` range its producer cannot
  meet, a network with no layer of tag 0, 1 or 2, a header `shift` or `gf` of 128 or more
  that no layer uses. Four checks the decoder made only when a network ran now come
  before the streams with the same code: a layer that does not fit the channels before
  it, a network that does not read the payload's channel count, a header `shift` or `gf`
  of 128 or more that a layer uses, and a leaky ReLU's shift of 64 or more. A convolution
  whose padded side exceeds `2^26` is refused.
- **The direct guard is test-only.** With this decode on its lowering, no production
  decode calls `int_hyper_synth::run_net_checked`, its per-layer checks or its IGDN check:
  the `check` ranges of every lowered network make them. They move into the test-only
  `int_hyper_synth::legacy` (feature `legacy_int`), beside the direct decodes that call
  them for the tests that hold the two to each other. `run_net` and `decode`, the
  unchecked path the integer world renderer uses, stay.
- **The refusal codes, written down.**
  [integer-payloads](extensions/integer-payloads.md) §14 said that every refusal of this
  capability is `payload_malformed`. The decoder reports `model_mismatch` for a bundle
  that breaks its rules or does not fit the payload, as it did before this revision and
  as its tests pin; §14 and integer-operators §4 now say so.

### Load step 8 in closed form

- **The bounds pass.** `wai_ops::bounds` now computes each `conv`'s and `conv_transpose`'s term hull and partial-sum interval ([integer-operators](extensions/integer-operators.md) §8.2) from a few sums per output channel, or per residue for `conv_transpose`:
  - the least and greatest weight;
  - the sums of the positive and of the negative weights. These are kept in `i64` where the tap count times the largest magnitude shows that every partial sum fits, and in `i128` otherwise.

  An `igdn` row now folds in `i128`. These forms give the rule's integers wherever no operation of its fold can saturate. The pass checks that condition: both input ends in `i128` with `l ≤ h`, the result within `±2^255`, and no `i128` overflow. Anywhere else it folds term by term in 512 bits, as before. §8.2 gains an informative note on the closed form. No certificate, verdict, refusal or output changes.
- **The oracle.** The pass as it stood is kept as `wai_ops::bounds_oracle`, test-only; another crate's tests reach it through the `oracle` feature. The two passes must agree node for node, with the same refusals and the same first failures, on:
  - every committed lowered graph and graph vector;
  - the pretrained conditional-codec set, when `WAI_SOTA_INT_DIR` names it;
  - random operators over input intervals the pass never forms (the ends of `i64` and `i128`, up to `2^255`, and `l > h`);
  - random graphs that reach saturation;
  - every graph the integer-operator differential's generators lower (`tests/int_ops_differential.rs`).
- **The lowering and the load.**
  - `Builder::finish` now returns the graph it parsed back from its bytes.
  - `whm1_net_loaded` and `wam1_net_loaded` (new) load that graph without writing and parsing its bytes a second time (`graph::load_lowered`, crate-only). They give `load`'s result, refusal for refusal. The `int_hyper`, splat and audio drivers use them.
  - The `WHM1` and `WAM1` lowerings now move each layer's lists into the graph instead of copying them.
  - The `WHG1` writer now adds a list's bytes at once and stores each value as one array.
  - The pass reads a node's weight range once.
- **SHA-256 on 64-bit Arm.** `wai-rs` now enables sha2's `asm` feature on `aarch64` targets other than Windows. The processor's SHA-2 instructions are used where it has them, detected at run time. The digests are the same, and the wasm32 build is byte-identical. The lockfiles of the workspaces that build `wai-rs` now include `sha2-asm`.
- **Cost.** For the pretrained conditional-codec set, lowering and loading both networks is now a small part of one decode, natively and in wasm32. A decode that loads its set now costs about what a decode that finds it held costs. The browser demo is one such decode: it decodes once per view.

### wai-quantum 0.6.0: the HIP, npu-cores, tensor-engine and avx512 members, the dataflow device's vector engine, and the gate as a portable tensor program

Breaking in three places (the last bullet). The default features, `full`, and every pinned
digest are unchanged.

- **The HIP member** (`quantum_fabric_hip`, opt-in, Linux): the CUDA member's kernel text,
  unchanged, compiled at run time by HIP's runtime compiler for each device's own
  architecture, with contraction off. The runtime and its compiler are loaded when first
  used; the device's properties are read in the runtime's frozen layout, or the device is
  refused. Its own implementation id (`hip-lanes64/…`), its own key (`gpu-hip-<arch>`); the
  mutants are the CUDA kernel's. With `quantum_sv`, the dense f64 state vector runs the CUDA
  kernels' text too (`hip-sv64/…`).
- **The npu-cores member** (`quantum_fabric_npu_cores`, opt-in, Linux): the gate's
  arithmetic on one compute core of a neural processor's tile array, programmed directly
  rather than through a model runtime. The core forms each gathered pair's four rows in
  exact 64-bit integers; the host checks their range and keeps the low 32 bits. The program
  (`npu_cores.cc` in `npu_cores_design.py`) is built at run time with the tile array's open
  toolchain and cached per variant, generation and toolchain; the host side
  (`npu_cores_host.cpp`) is compiled against the device runtime and loaded. One program
  serves every gate and the arithmetic probe; each mutant is its own program. Key
  `npu-cores-<pci vendor>-<device>`.
- **The tensor-engine member** (`quantum_fabric_tensor_engine`, opt-in): the gate's sixteen
  products per pair as exact limb convolutions in one bf16 matrix multiply per tile of pairs,
  accumulated in fp32, on a tensor processor's matrix engine. Each value is four 8-bit limbs,
  which bf16 holds exactly; every limb product is at most `2^16` and every coefficient a sum of
  at most four, so fp32 holds every partial sum exactly in any order, on any engine that
  multiplies bf16 and accumulates in fp32. The host checks each coefficient is an integer
  below `2^24`, assembles the products and does the rest. The kernel runs as a helper under
  the engine's kernel toolchain; its target here is that toolchain's simulator, and every
  certificate says so. One mutant is nine-bit limbs, which the engine's inputs cannot hold.
- **The avx512 vector unit** (`quantum_fabric_vector`): eight 64-bit lanes, each `fx` an
  exact 32×32 multiply, an add and the unit's own 64-bit arithmetic shift. Discovery opens
  one member per instruction set the processor reports, so a processor with both units has
  both members.
- **Measured** on an integrated GPU (gfx1151), a neural processor's tile array and the
  processor beside them, all silicon: the HIP member 9/9 (66020 probe pairs, all four pinned
  certificates, circuits to 22 qubits, 6 of 6 mutants) and its state vector 3/3 (the pin
  `b7d129d0…`, 4 of 4 mutants); the npu-cores member 9/9 (66020 probe pairs, all four
  certificates, circuits to 20 qubits, 3 of 3 mutants); avx2 and avx512 9/9 each; the
  tensor-engine member on its simulator 9/9 (66020 probe pairs, all four certificates, 4 of 4
  mutants); and,
  unchanged, the WebGPU member (`gpu-vulkan-1002`) and the OpenCL member
  (`opencl-gpu-1002`, the state vector's pin included) bit-exact there. CI runs each of them
  on that host.

- **The vector engine as a member** (`quantum_fabric_dataflow`, key `dataflow-1e52-b140-vector`):
  on the device generation whose vector engine multiplies 23-bit numbers, the gate's
  arithmetic runs on that engine, 32 lanes at a time, on 32 compute tiles. Each tile's
  first core configures the unpacker and the packer and hands the tile its instructions
  inline; the unpacker copies the pairs' columns into the destination registers as
  32-bit integers, the vector engine forms the rows, the packer copies them back. Each
  32×32 product comes from four 23-bit partial products, corrected for sign; `fx` is
  held as 30 low bits and the rest, and each row's terms are added or subtracted with a
  carry after each. A pass is sized to its share of the pairs, in steps of 64.
- **`quantum_fabric::tensix`**: the compute tile's instructions the programs use, their
  opcodes and fields as the device simulator's own table gives them; CI checks the
  transcription against that table. **`quantum_fabric::vector_engine`** builds the
  programs, the probe included; the assembler check covers them too.
- **Measured** on the device simulator, built from source: 66020 probe pairs agree, all
  four pinned certificates, circuits up to 20 qubits match native, 3 of 3 kernel mutants
  (the same three the other members carry) killed. Every certificate is labelled
  simulator.
- **The gate as a portable tensor program** (`quantum_fabric::stablehlo`): the gate's rows
  and the arithmetic probe as a StableHLO module, the input of compilers for tensor
  processors, GPUs and CPUs. Every value is a 64-bit integer that never overflows
  (`|e·a| < 2^62`; a row of four `fx` stays below `2^35`), so a compiler must give the
  exact rows whatever it does with them. It is not a member: the program leaves the
  crate, a compiler's runtime runs it, and its outputs are compared with the exact ones. The
  `stablehlo_gate` example writes the module, its three mutants (the same three), a
  batch of 4096 pairs whose gate and states reach the 32-bit edges, and the outputs a
  build must give.
- **Measured** with two independent open-source tensor compilers, IREE 3.12.0 and XLA
  (through JAX 0.11.2, the compiler tensor processors take this program through): exact
  on the x86-64 CPU with each; on every Vulkan device of the hardware host (a discrete
  GPU, an integrated GPU, and a software one); on the discrete GPU through CUDA with
  each; through Metal on two generations of the macOS platform's GPU; with XLA on the
  arm64 CPU. On the CPU each mutant fails where it should under each compiler; on every
  other device, the third mutant's rows do. Builds for three GPU
  generations the fleet lacks compile but are not run. `ci/tensor-program/check.sh`
  holds each run to its outcome and fails a run that ends without stating one, so a
  device that cannot run reads as a failure, never as a mutant caught. CI:
  `wai-quantum-tensor-program`. That host's GPU runs it through HIP, built for the GPU's own
  architecture, and on both of its Vulkan devices.
- **The products on a matrix unit, as a portable program** (`stablehlo::limb_module`): the
  tensor-engine member's limb layout as one StableHLO `dot_general` of bf16 operands
  accumulated in f32, the program a compiler maps to a matrix unit. With it, the example writes
  a nine-bit-limb data set whose exact coefficients bf16 cannot reproduce, so a run that
  matches them is not comparing. Exact under both compilers on the x86-64 and arm64 CPUs and
  on the discrete GPU through CUDA. Not exact through one compiler's HIP build for the
  integrated GPU, whose bf16 multiply returns values that are not integers; CI shows that
  step and does not require it.
- **Breaking:**
  - an OpenCL key is `opencl-<kind>-<vendor id>`, the device's `CL_DEVICE_VENDOR_ID` in hex
    (`opencl-cpu-8086`, `opencl-gpu-10de`, `opencl-gpu-1027f00`, `opencl-gpu-1002`), not a word
    of the vendor's name; `opencl::DeviceIdentity` gains `vendor_id`, which the member id
    binds, and `vendor_slug` is gone;
  - `quantum_fabric_vector::Isa` gains `Avx512` (`Isa::ALL` has four), and `native_isas` lists
    every set written for the architecture;
  - `quantum_fabric_dataflow::Chip` gains `vector`.

### wai-quantum-qpu 0.2.0

Builds against wai-quantum 0.6.0; nothing of its own changes.

### Parameter sets kept loaded between decodes

- **The store.** A decode through the integer operator set lowers its parameter set and
  loads each graph, the bounds pass included; for the pretrained conditional-codec set
  that costs more than the decode. `int_set_cache::LoadedSets` now keeps a set once it
  has loaded, and the next decode through the same bytes takes it from there
  ([integer-operators](extensions/integer-operators.md) §8.6, informative note).
  `int_payload::decode_int` and `int_payload::declared_cost` use it for `WIH1`, a
  `WIV1`'s keyframe, `WIA1`, `WIS1` and `wai.neural.int_synth`, and so the neural sink
  and `wai-web`'s `decodeWiv1`, `decodeWih1`, `decodeWia1`, `decodeWis1` and
  `decodeIntSynth` do too; `decodeIntSynth` now decodes through `decode_int`, and a
  refusal's message starts with its code. `int_synth::decode_from_spec`, which the
  byte-equality crate shares, is now three steps (`check_payload`, `load_spec`,
  `decode_loaded`) and loads its set every time. The entry points that read the tools'
  JSON forms do not use the store.
- **The key** is the capability, the lowering, the profile the graphs were loaded under,
  and the SHA-256 of the listing of the members the decode reads, each by its own
  SHA-256 (`pin::listing_from`): the digest a pin over exactly those members names.
  Never a name: two sets whose members share their names and differ in one byte have
  two entries. Parameters `int_payload::resolve` built carry the listing digest it
  already formed for check 5, so such a decode does not hash its set a second time;
  the digest is used only for a capability that reads exactly the members it was formed
  over, and any other key hashes the bytes.
- **What does not change.** An entry is made only from what a load returned, after it
  returned, so a refused set is never kept: it is lowered, loaded and refused again on
  its next decode, with the same refusal. Every check before the load runs as before,
  in integer-payloads §7's order. Every output, refusal and declared cost is unchanged,
  and no golden moved.
- **Memory.** A kept set is the set's, outside every decode's declared working bytes
  ([integer-payloads](extensions/integer-payloads.md) §7, informative note). The store
  keeps its entries within a byte budget, 256 MiB to start with, evicts the least
  recently used first, and keeps no set larger than the whole budget; a budget of 0
  turns it off, and the set is then not hashed. `LoadedSets::set_budget` sets it, and
  `wai-web` exports `setLoadedSetBudget` and `loadedSetStats` for a page. An entry is
  counted as the shared allocation that holds the set, its graphs' heap bytes
  (`wai_ops::Graph::heap_bytes`, new) and a fixed share of the store's indexes, never
  less than it holds (`tests/int_payload_cost.rs`).
- **Tests.** `tests/int_set_cache.rs`: a second decode through each committed set is a
  hit, by the store's count, with the bytes of a decode with the store off; sets one
  byte apart never share an entry; a set the load refuses (an `int_synth` set the bounds
  pass refuses; a conditional-codec set whose lowering admits no input range) stays
  refused while a set one byte away is held; every refusal fixture keeps its code and
  detail with its set held; the least recently used set goes first; threads decoding
  one set keep one entry. The browser crate's check (`demo/verify_int_payload_wasm.mjs`) decodes `tiny`
  twice and counts the hit.

### `wai.neural.int_hyper` and `wai.video.int_hyper` through the integer operator set

- **The decode.** `int_hyper_synth::decode_conditional_checked` (so the `WIH1` dispatch,
  the keyframe of a `WIV1` with keyframe 1, the staged sink and the browser crate) plans
  both networks as [integer-payloads](extensions/integer-payloads.md) §7 orders, checks
  the payload's shapes against them, lowers `h_s` by `whm1/1` and `g_s` by `whm1/1` with
  the RGB8 step ([integer-operators](extensions/integer-operators.md) §13.2, §13.0),
  loads each under `acc64` with its bounds certificate, decodes the hyper-latent, runs
  `h_s` through the operator set's interpreter, splits its output into `σ` and `μ`,
  decodes the latent against the tables `σ` selects, and runs `g_s`. Every
  convolution, transposed convolution and inverse GDN runs on the fast kernels behind
  the interpreter. The graphs' `check` ranges are the checks the direct decode made
  before each layer, so the same streams are refused (`field_range`). Every committed
  output is unchanged, byte for byte (the conformance vector, the neural-keyframe clip,
  and the pretrained set's image and clip regenerated in `wai-sota-vectors.yml`), and
  every committed refusal fixture keeps its code. The direct decode is kept, test-only
  (feature `legacy_int`), in the differential tests and their nightly run.
- **Planned breaks** (SPEC §8, integer-operators §15.3 item 3), as for
  `wai.splat.int_codec`, in either network: new refusals of sets no committed corpus
  contains, each `model_mismatch` after the payload's shapes are checked and before
  either stream is decoded. A scan of every committed set of the two capabilities, and
  of every committed `WIH1` payload wherever it is carried, finds none affected.
- **The inverse GDN on the fast kernel.** The interpreter now lends an inverse GDN to the
  fast IGDN kernel, as it lends convolutions, wherever the values present keep every
  term and norm within `i128` and the root and output within `i64`, which every
  certified graph does; the splat and audio decodes use it too. Their outputs are
  unchanged.
- **The unchecked decode is removed.** `int_hyper_synth::decode_conditional`, which ran
  both networks with no exactness check, is gone; its callers (the examples, the
  benchmark, the tests and the byte-exact neural-keyframe golden) decode through
  `decode_conditional_checked`, so through the lowering with its `check` ranges, with
  every output equal. The layer-by-layer guard `run_net_checked` stays while
  `wai.neural.int_mlicpp` still runs its networks through it.
- **The tools' JSON models.** The browser crate's debug decode of the tools' JSON
  form now refuses (`model_mismatch`) a JSON model holding a weight past `i16` or an
  inverse-GDN `γ` past `i32`, the widths `WHM1` gives them and its lowering writes;
  the committed models hold none.
- **Their `ops` member in the registry** ([integer-operators](extensions/integer-operators.md)
  §11). `registry/capabilities.json` gives `wai.neural.int_hyper` and
  `wai.video.int_hyper` the member `{"set": "wai-ops/1", "profile": "acc64", "subset":
  "full", "lowering": "whm1/1", "stage": {"rans64": {"table": "t4097", "escape":
  "lenient", "end": "canonical"}, "bucket": "ceil"}}`. SPEC §5's rows cite the payloads
  ([integer-payloads](extensions/integer-payloads.md) §4, and §3 for the clip) and the
  lowering (integer-operators §13.2), and SPEC §7 names both capabilities among those
  that decode through the operator set.

### wai-quantum 0.5.1: neural-processor, matrix-extension, vector-unit, signal-processor and dataflow members

Additive. The default features, `full`, and every pinned digest are unchanged.

- **The neural-processor member** (`quantum_fabric_npu`, opt-in, macOS only). A neural
  processor computes in fp16 and does not document how it accumulates, so the member
  never relies on it: each 32-bit operand is split into eight 4-bit limbs, and the
  processor computes the fifteen coefficients `Σ_{i+j=s} a_i·b_j` of each product. Every
  limb product is at most 225 in size and every coefficient at most 1800, so every value
  it forms is an integer fp16 holds exactly, in any order. The host assembles each
  product exactly and applies `lanes::apply_words`'s arithmetic.
- **Credited by placement.** The member writes its model (one multiply, one sum) as the
  platform runtime's model description, compiles it at run time, and reads the
  runtime's compute plan: every operation must be placed on the neural processor, or
  the member is refused with the placement reported, so no CPU result is credited to the
  processor. It is then probed and certified against the pinned corpus digests, and
  the two model mutants (the multiply made an add, the sum a maximum) are caught.
- **The same program, portable.** `npu::onnx_limb_model` writes it in the portable model
  format (ONNX, opset 17: one `Mul`, one `ReduceSum`), with fp16 or int32 tensors, and
  `npu::expand_i32` lays out the int32 inputs; `examples/npu_onnx.rs` writes both models,
  a batch of inputs and the exact products they must give, so a neural processor
  reached through a runtime that reads that format can be checked against them.
- **Fixed** before release: `npu::f16_of(±2048)` subtracted past zero in a debug build.
- **The matrix-extension member** (`quantum_fabric_sme`, opt-in, aarch64 only): the
  gate's fixed-point arithmetic in the processor's streaming vector mode (SME), 64-bit
  lanes, by inline assembly; the host gathers each gate's pairs and checks each row's
  range. Its instructions exist only in that mode, so a kernel that ran, ran there.
  Detection reads the processor's feature flags (on Linux, `AT_HWCAP2`, which a
  user-mode emulator supplies for its guest); a runner under an emulator declares it,
  and every certificate carries that provenance. Three kernel mutants are compiled in.
- **The DSP vector-unit member** (`quantum_fabric_dsp`, opt-in, in a build for the signal
  processor with its 128-byte vector extension). The gate's arithmetic in 32-bit lanes,
  each 64-bit value a pair of words: the product is the vector unit's exact 32×32
  multiply (the 32-bit operand times the other's low half, unsigned, plus its high
  half, signed, sixteen bits up); the rounding constant's carry and each row's carries
  and borrows go into the high word; the shift is arithmetic over all 64 bits. The
  fabric gains the class `dsp`. Three kernel mutants are compiled in.
- **The CPU vector-unit members** (`quantum_fabric_vector`, opt-in): the gate's arithmetic
  in the processor's own vector extension — `avx2` on x86-64 (the 64-bit arithmetic
  shift, which the unit lacks, built from a logical shift and the sign bit), `sve` on
  aarch64 and `rvv` on riscv64, both written once for every vector length (the length
  set per pass and read back from the unit into the member's id). Silicon for `avx2`;
  `sve` at 128, 512 and 2048 bits and `rvv` at 128 and 1024 bits under the user-mode
  emulator. Three kernel mutants per instruction set.
- **The dataflow member** (`quantum_fabric_dataflow`, opt-in, Linux): the gate's arithmetic on
  the first RISC-V core of each of 32 compute tiles of a dataflow accelerator, each over a
  contiguous slice of a gate's pairs, two device generations (PCI 1e52:b140 and 1e52:401e),
  driven through the device's own library at the level of configuration space, memory
  windows, reset registers and clocks. The program comes from a new RV32IM
  encoder (`quantum_fabric::rv32`), whose listings an independent assembler turns into
  the same words; every 64-bit value is a pair of words. The library is the device's
  open-source simulator, built from source, and every certificate is labelled simulator.
- **The signal processor as a processor.** The bit-identity digests, the frame fixtures and
  the byte-pinning library tests pass for the processor's Linux target under its
  user-mode emulator. Emulator releases before 11.0 test a whole predicate register
  where the processor tests its low bit, in a branch on a predicate written in the same
  packet; compiled code relies on the low bit, and one library test fails under such an
  emulator. `ci/dsp/predicate_branch.c` shows which emulator does which, and the
  suites run on 11.1.1, built from its release.
- **CI.** `wai-quantum-npu.yml`: the device tests on the macOS runner's neural processor,
  and the same build on Linux, where the member is absent and says so.
  `wai-quantum-sme.yml`: the matrix-extension member's device tests for aarch64 Linux
  under the host's user-mode emulator, certificates labelled emulator; and the macOS
  runner, whose processor has no SME, finding no member. `wai-quantum-dsp.yml`: the
  signal processor's suites and the vector-unit member's device tests under the pinned
  emulator, after the emulator's own branch probe, and the same build on x86_64, where
  the member is absent and says so. `wai-quantum-vector.yml`: `avx2` on the container
  runner's processor, `sve` and `rvv` under the host's emulator at the lengths above (each
  leg checks the length the member read), and the macOS runner finding no member.

### `wai.audio.int_codec` through the integer operator set

- **The decode.** `int_audio::decode_typed` (so the `WIA1` dispatch, the staged sink and
  the browser crate) plans the synthesis as
  [integer-payloads](extensions/integer-payloads.md) §7 orders, lowers `model.wam` by
  `wam1/1` ([integer-operators](extensions/integer-operators.md) §13.3, with its PCM
  step), loads the graph under `acc64` with its bounds certificate, decodes the stream,
  and runs the graph through the operator set's interpreter with `t_out` as its scalar
  and the fast transposed-convolution kernel behind it. Its `check` ranges are the
  checks the direct decode made before each layer, and the PCM step's range is the one
  in which a sample's scaling fits `i64`, so the same clips are refused
  (`field_range`). Every committed output is unchanged, byte for byte, and every
  committed refusal fixture keeps its code. The direct decode is kept, test-only
  (feature `legacy_int`), in the differential tests and their nightly run.
- **Planned breaks** (SPEC §8, integer-operators §15.3 item 3), as for
  `wai.splat.int_codec`: new refusals of sets no committed corpus contains, each
  `model_mismatch` before any stream is decoded, the PCM step's range among the `check`
  ranges whose producer must meet them. A scan of every committed audio set and payload
  finds none affected.
- **Its `ops` member in the registry** ([integer-operators](extensions/integer-operators.md)
  §11). `registry/capabilities.json` gives `wai.audio.int_codec` the member
  `{"set": "wai-ops/1", "profile": "acc64", "subset": "full", "lowering": "wam1/1",
  "stage": {"rans64": {"table": "t4097", "escape": "lenient", "end": "canonical"}}}`.
  SPEC §5's row cites the payload ([integer-payloads](extensions/integer-payloads.md) §5)
  and the lowering (integer-operators §13.3), and SPEC §7 names the capability among
  those that decode through the operator set.

### `wai.splat.int_codec` through the integer operator set

- **The decode.** `int_splat::decode_typed` (so the `WIS1` dispatch, the staged sink and
  the browser crate) plans the networks as
  [integer-payloads](extensions/integer-payloads.md) §7 orders, lowers `g_s.whm` by
  `whm1/1` ([integer-operators](extensions/integer-operators.md) §13.2, with the RGB8
  step of §13.0), loads the graph under `acc64` with its bounds certificate, decodes the
  stream, and runs the graph through the operator set's interpreter with the fast
  transposed-convolution kernel behind it. Its `check` ranges are the checks the direct
  decode made before each layer, so the same latents are refused (`field_range`).
  Every committed output is unchanged, byte for byte, and every committed refusal
  fixture keeps its code. The direct decode is kept, test-only (feature
  `legacy_int`): `tests/int_ops_differential.rs` decodes random sets and clouds both
  ways, every push and for an hour per build profile each night.
- **Planned breaks** (SPEC §8, integer-operators §15.3 item 3): new refusals of sets no
  committed corpus contains, each `model_mismatch` before any stream is decoded. A layer
  with a bias of magnitude `2^63 − 1` or more, or a `check` range its producer cannot
  meet, refused every decode with `field_range` when the network ran; a network with no
  layer of tag 0, 1 or 2, or with a header `shift` or `gf` of 128 or more that no layer
  uses, decoded where the shapes allowed. A scan of every committed splat set and
  payload finds none affected.
- **The interpreter's memory.** A run releases each tensor once its last reader has
  run, takes its inputs by value, and holds a transposed convolution's uncropped plane
  in `i64` (each cell is within the certified `convT.acc`). With its set loaded, a splat
  decode through the operator set holds no more than the payload's declared working
  bytes (integer-payloads §7, *Memory and work*); the declared cost of a `WIS1` payload
  is unchanged. `int_payload::declared_cost` now refuses, with the decode's code, a set
  whose lowering does not load.
- **The byte-equality matrix** (`wai-byte-equality.yml`, `byte-exact-conformance`) now
  decodes the committed `WIA1` and `WIS1` payloads of `wai.audio.int_codec` and
  `wai.splat.int_codec` on every leg, under the reference and the fast kernels, to their
  pinned PCM and attribute digests; it compiled neither decoder before.
- **Its `ops` member in the registry** ([integer-operators](extensions/integer-operators.md)
  §11). `registry/capabilities.json` gives `wai.splat.int_codec` the member
  `{"set": "wai-ops/1", "profile": "acc64", "subset": "full", "lowering": "whm1/1",
  "stage": {"rans64": {"table": "t4097", "escape": "lenient", "end": "canonical"}}}`.
  SPEC §5's row cites the payload ([integer-payloads](extensions/integer-payloads.md) §6)
  and the lowering (integer-operators §13.2), and SPEC §7 names the capability among
  those that decode through the operator set.

### wai-quantum 0.5.0: provenance, sampling members, device programs, and three more members

Breaking: `Certificate` gains a public field, `provenance`, and `FabricClass` is now
`#[non_exhaustive]`, so a class can be added in a minor release. A silicon certificate
hashes exactly as before; every pinned digest is unchanged. The `quantum_fabric` feature now
includes `quantum_toolchain`.

- **Provenance in the certificate.** `Provenance::{Silicon, Emulator(name), Simulator(name)}`,
  inside the certificate's hash unless it is silicon. Placement ranks silicon first, then
  emulators, then simulators, so a simulated member never stands in for hardware that is
  present.
- **Sampling members.** The kernels `wai.quantum.circuit/sample` and
  `wai.quantum.circuit2/sample`, the `Sampler` trait, `certify_sampler` (classical fidelity
  against the exact distribution, over stated shots) and the `qpu-v1` corpus. `VirtualQpu`
  samples a stated noise model (global depolarizing and readout flips) and says it is a
  simulator. `VirtualDevice` adds a coupling map and a native basis: each circuit is
  compiled, the program verified, and the gate errors charged for the program it would
  run, routing included.
- **Programs for devices** (`quantum_toolchain`). `device::device_qasm` writes OpenQASM 2
  or 3 in the dialect's standard gates only, with every multi-controlled gate lowered
  exactly (dyadic angles throughout; refused, never rounded, past the table) and every
  qubit measured into `c[i]`; `counts_from_bitstrings` and `counts_from_integers` read a
  service's counts back. `native::compile` targets a processor's own basis (`rz`/`sx`/`x`/`cz`,
  `prx`/`cz`, or `gpi`/`gpi2`/`zz` in turns) on its own qubits, routed over its coupling
  map; `NativeProgram::verify` checks a program against its circuit by simulation before
  it is sent. `from_qasm` now reads `ch`.
- **The OpenCL member** (`quantum_fabric_opencl`, opt-in): the exact 64-bit kernel in
  OpenCL C, built by each device's own compiler; the ICD loader is opened when first used.
  Certified bit-exact on GPUs and CPUs of several vendors; an OpenCL accelerator is refused,
  since no fabric class names it.
- **The FPGA member, simulated** (`quantum_fabric_fpga`, opt-in): the datapath in Verilog,
  each 32×32 product from four 16-bit partial products, compiled at run time by Verilator
  into a cycle-accurate simulation. Its certificates say they were earned in a simulator.
- **The dense f64 state vector on accelerators** (`quantum_fabric::sv64`, with
  `quantum_sv`): the CPU's own launches (diagonal with the exact-one skip, the one-qubit
  formula, and ordered row sums), each an IEEE binary64 operation with no fused
  multiply-add, on the CUDA member (`CudaMember::sv`) and on OpenCL devices with binary64
  (`OpenClMember::sv`). Certified against the pinned corpus digest `SV_DEFAULT_PIN`; a
  device without binary64 refuses, saying so.
- **wai-quantum-qpu 0.1.0** (new crate): clients for three quantum processing protocols
  (`sampler_v2`, `prx_circuits`, `native_turns`), each compiling exactly to the device,
  verifying before submission and reading counts back onto the circuit's qubits;
  `QpuSampler` makes a service a fabric member; the `wai-qpu` command reads a device,
  runs the bit-order probe, and certifies. `curl` is the transport, with secrets kept off
  command lines. Tested against services that follow each protocol's published contract.
- **CI.** Workflows for the OpenCL member, the simulated FPGA member and the QPU clients;
  the CUDA and OpenCL hardware jobs certify the f64 state vector; the hardware host
  declares its devices in its runner's environment.

### wai-quantum 0.4.4: the fabric's CUDA member

Additive. The default features, `full`, and every pinned digest are unchanged.

- **The CUDA member** (`quantum_fabric_cuda`, opt-in, native targets only): the gate
  kernel in CUDA C with exact 64-bit integers, compiled at run time by the runtime
  compiler for each device's own architecture (`--fmad=false`), one launch per gate. The
  driver and the compiler are loaded when first used, so a build with the feature runs on
  a host without CUDA and finds no member there, saying why. Each device is probed against
  the CPU's arithmetic and certified bit-exact by reproducing the pinned corpus digests; a
  result that would leave 32 bits is refused, never returned.
- **CI.** `wai-quantum-cuda.yml`: the device tests on a hardware host's GPU (every
  certificate, every kernel mutant caught), and the same build on a host with no CUDA.

### Reader limits for the extension formats' durations

A score, haptic signal, audio scene, splat4d clip or film declares its length in its
contract, and rendering it allocates in proportion. Nothing bounded the declaration
(found while fuzzing: a score's `mixdown` allocated `duration` samples outright).

- **The limits** (each extension's §2.1, `wai::limits`), shared with SPEC §2:
  - score and haptics: `duration × channels` at most 2²⁷ samples;
  - audio scene: `duration` at most 2²⁶, so its stereo floor is at most 2²⁷;
  - splat4d (`n_frames`), film (`frames`) and a reel's shots together: at most 2²⁰
    frames.

  A reader may refuse past a limit (`duration_too_long`), when it reads the contract.
- **Film tick arithmetic.** `start_tick + f × ticks_per_frame` wrapped in release
  builds and panicked in debug ones. A reader now refuses a schedule whose last tick
  passes 2⁶⁴ − 1, and `frame_tick` saturates rather than wraps.
- **Tests.** `tests/extension_duration_limits.rs` rewrites a committed vector of each
  format to sit exactly at its limit (it reads) and one past it (refused by name), and
  checks the film schedule at and past 2⁶⁴ − 1. The old readers fail all three tests,
  and every committed corpus verifies unchanged.

### Coverage-guided fuzzing of the reader

`wai-rs/fuzz/` adds four libFuzzer targets, built by cargo-fuzz with AddressSanitizer:
- `envelope`: the `WAI1` and `WAI2` readers and the listing form;
- `int_payload`: every integer capability, against the committed parameter sets;
- `media`: the decoders of the floor and the recommended set;
- `containers`: the extension containers.

They are seeded from the repository's own vectors. `wai-fuzz.yml` runs each target for
15 minutes every night and keeps the corpus between runs; it never runs on a pull
request. `fuzz/run.sh` runs the same targets anywhere a nightly toolchain is installed.
SECURITY.md now says what is fuzzed and what is not.

- **Found and fixed.** In its first seconds, `containers` found the audio scene reader
  reserving entries for a `u32` count read from the input, before it read one: 48 GiB
  asked for by a few bytes. The world, reel, score, haptic and splat4d readers did the
  same. Each now reserves no more than its remaining bytes could hold.
  `tests/container_reservations.rs` rewrites a committed vector of each to declare
  2³² − 1 entries in four bytes and requires a refusal with no allocation past a few
  megabytes; on the old readers the world asset table alone asked for 256 GiB.

### Reader limits for picture, audio and video payloads

The last decoders that allocated what a payload declared now hold it to limits
(SPEC §2), as manifests and text payloads already were.

- **The limits.** A reader must accept a picture up to 65 536 pixels on a side and
  2²⁷ pixels, audio up to 2²⁷ samples counting every channel, and AV1 up to 2²⁰
  temporal units of such pictures. It may refuse past them (`image_too_large`,
  `audio_decoded_too_long`, `video_too_large`), from what the payload declares wherever
  it declares it. The writers refuse the same.
- **Payload framing** (SPEC §5). The AV1 framing that the reference has always used,
  and packaging §7 writes, is now in the standard itself. So are the FLAC and Ogg Opus
  forms, and the rule that an image payload decodes only as its own format.
- **Defects fixed in the reference decoders.**
  - A FLAC `STREAMINFO` declaring 2³⁶ − 1 samples made the decoder ask for 256 GiB up
    front.
  - A 24-bit FLAC decoded to about ±256 instead of ±1.
  - An Ogg Opus stream whose source rate was not an Opus rate (44.1 kHz) was refused.
  - `png_decode` and `jpeg_decode` decoded any format the bytes resembled.
  - `jxl_decode` refused a deeper-than-8-bit picture and returned one or four channels
    where it promised three.
  - The AV1 decoder decoded every packet before truncating to `n`. It panicked on
    monochrome, misread 4:2:2, 4:4:4 and 10- and 12-bit frames, and let dav1d allocate
    whatever a sequence header declared.
- **Streaming AV1.** `av1_decode_each` hands on one frame at a time. `av1_decode` and
  `wai_video_av1_decode`, which return every frame at once, refuse past 4 GiB of output
  from the header (`video_decoded_too_large`). `wai decode` streams frames to its file.
- **Unchanged.** Every 8-bit 4:2:0 frame decodes byte for byte as before. The AVIF path
  now calls libavif directly with the reader limits as its decoder limits (it drops
  `libavif-image` and the rav1e build inside libavif), and `include/wai.h` is
  regenerated.
- **Tests.** `tests/media_decode_bounds.rs` checks each limit at its edge and one past
  it, and each defect above. CI runs it in `codec-wrappers`. Five mutations, each
  reintroducing one defect, fail it.

### Vector graphics (Draft): a title card and credit roll in the corpus

`wai/vector-conformance/` gains `motion/stroke-titles`: a 1280 × 720 document at 25 fps
in the default stroke set, a lower third that slides in, holds and fades as a group, and
a credit roll that rises three design pixels a frame. Each line's window is cut to the
frames at which it meets the canvas (graphics-vector §5.4, *Where a block is*). Every
frame's digest is in its expected file, so the native, WASI, Node and three-engine
browser legs and the independent checker each check all of them.

### `wai.neural.int_synth` through the integer operator set

The first capability whose reference decoder decodes through `wai-ops/1`.

- **The decode.** `int_synth::decode_bytes` checks the payload
  ([integer-payloads](extensions/integer-payloads.md) §13), lowers the set by
  `int_synth/1` ([integer-operators](extensions/integer-operators.md) §13.1), loads the
  graph under `acc64` (load steps 1–8, the last being the bounds certificate) and runs it
  through the operator set's interpreter, with the fast transposed-convolution kernel
  behind it (`wai-rs/src/int_ops.rs`). Every committed output is unchanged, byte for
  byte, on every platform the byte-equality matrix covers and in the browser crate, and
  every committed refusal keeps its code (`payload_malformed`). The direct decode it
  replaced is kept, test-only (feature `legacy_int`): `tests/int_ops_differential.rs`
  decodes random sets and payloads both ways and requires the same output or both
  refusals outside the planned breaks, and `wai-ops-differential.yml` runs it for an
  hour in each build profile every night.
- **Planned breaks** (SPEC §8, integer-operators §15.3): new refusals of input no
  committed corpus contains. `wai.neural.int_synth` refuses a payload whose reserved byte
  is not 0, whose `C` is 0 or whose latents exceed `2^26` values; a set with no layers,
  with a layer whose `out_ch`, `k` or `stride` is 0, with a field of `2^32` or more, or
  with a shift of 128 or more; a set or payload a cap refuses (a tensor, or a transposed
  convolution's uncropped output, past `2^26` elements); and a set whose certificate does
  not meet `acc64`. A shift from 64 to 127, which a build used to mask or panic on, now
  has its exact meaning. A scan of every committed `int_synth` set and payload finds
  none affected. The text now states the two rules the lowering already made and the
  payload text did not: at least one layer, and `out_ch ≥ 1`.
- **Its `ops` member in the registry** ([integer-operators](extensions/integer-operators.md)
  §11). `registry/capabilities.json` gives `wai.neural.int_synth` the member
  `{"set": "wai-ops/1", "profile": "acc64", "subset": "full", "lowering":
  "int_synth/1"}`, and the schema describes the member. SPEC §5's row cites the payload
  and set ([integer-payloads](extensions/integer-payloads.md) §13) and the lowering
  (integer-operators §13.1), and SPEC §7 says that a `NeuralIntegerExact` capability
  that decodes through the operator set names it in its `ops` member.

### Vector graphics (Draft): the sink as a WebAssembly 1.0 module, in three browser engines

Additive. No payload, outline set, vector or digest changes.

- **The sink's exports.** `wai-vector/wasm` exports the sink of
  [graphics-vector](extensions/graphics-vector.md) behind a C ABI, with no bindings
  generator and no imports. `wv_add_set` holds an outline set a payload may name;
  `wv_open` loads a payload to a handle, or to its refusal's code; `wv_render` gives a
  frame's `RGBA8P` at a listed size and rate, or the request error's code; `wv_cost`
  gives a listed size's cost; `wv_transcript` gives the transcript, and `wv_projection`
  the projection at a frame of a listed rate, each as the JSON of §5.4, or the request
  error's code; `wv_sha256` gives a digest; `wv_close` releases a handle. A result is
  read at `wv_out_ptr`, `wv_out_len` bytes of it. No export takes or returns a 64-bit
  integer: a frame number goes in as two halves.
- **The package.** `wai-vector/wasm/build-pkg.sh` builds the module a site serves: the
  sink's exports alone, with no glue script. It is WebAssembly 1.0 and nothing later:
  no SIMD, no threads, no bulk memory, no sign-extension operators, no reference types.
  The target is `wasm32v1-none`, whose precompiled `core` and `alloc` are 1.0
  themselves, so the crate is now `no_std` and carries its own allocator. The build
  then reads the module back: `wasm/check-mvp.mjs` decodes every section, type, segment
  and instruction, and refuses anything later than 1.0, any import, any 64-bit integer
  at an exported function, any floating point, and any exported function the script
  does not list.
- **Three browser engines.** `wai-vector/browser/run.mjs` serves a page, the package and
  the corpus, and in each engine loads every payload of `vector-conformance` to its
  expected code, checks its SHA-256 and each listed size's cost, and renders its frames
  through the page's hook, `window.__wai_overlay.frame(size, rate, g)`, which returns
  the renderer's own buffer: every listed frame at sizes up to 1280 × 720, and sampled
  frames above. The page hashes each frame with the module's SHA-256 and with the
  platform's digest. The sampled frames are also sent to Node as bytes, and Node hashes
  them itself. Where a payload's expected file states a text record, the page's hook
  also gives the sink's transcript and its projection at every frame of every listed
  rate, `transcript()` and `projection(rate, g)`; written out as the record's lines,
  they must be the committed `<name>.a11y.jsonl` byte for byte. The recomputation
  module then recomputes every expected file and authoring form in each engine.
  Nothing is read from a canvas. A run passes only when all three engines ran: a
  missing engine is a failure.
- **CI.** `wai-vector.yml` gains the `browser` job: the package built and read back,
  then the three engines, in a container image pinned by digest.

### Vector graphics (Draft): text and accessibility, §5 of `wai.graphics.vector`

- **[graphics-vector](extensions/graphics-vector.md) §5 is written.** A payload carries
  its words beside its glyphs, and this revision says what a sink computes from them and
  what it owes the viewer:
  - **The projection and the transcript, byte for byte (§5.4).** A block's active set is
    the union of the effective windows of the runs and paths that draw it, cut to the
    document's window, sorted, with windows that overlap or touch merged. The projection
    at a frame is every block active at the frame's tick, in reading order; the
    transcript is every block with its active set. Both are JSON with a fixed member
    order and exactly three string escapes. A block's `purpose` is written as the byte
    carried, an unregistered code included; `mode` is in every node; a ruby link is
    written as the base block's index with the range's start and length; `decorative`
    blocks are in neither. Opacity, position and clips do not enter either value. A
    payload's *text record* is its transcript and every frame's projection at every
    listed rate, one value a line: text-equivalence is byte equality of it.
  - **Encoder rules (§5.1, §5.3)** that a verifier checks and a sink does not: one base
    direction for a block, one joining sequence in one run, vertical text and its
    `mode`, Normalization Form C, every byte of a string in exactly one cluster, no
    zero-glyph cluster for a visible character, and no colour glyphs.
  - **The purpose vocabulary (§5.2),** with how each purpose is exposed and whether it is
    announced by default.
  - **Exposure (§5.5).** A presenting sink exposes the current projection to the
    platform's accessibility interface, updates it when the projection changes, inserts
    strings as text and never as markup, and keeps within stated announcement limits.
    What assistive technology then says is not claimed.
  - **Whether the words match the glyphs (§5.6):** the chain from a parent font to an
    outline set, from a text block to glyph IDs, and from glyphs to the overlay layer;
    what an authoring receipt records; and what the chain does not reach.
  - **Captions under the viewer's control (§5.7).** Captions travel as timed text. A
    sink that cannot read the platform's caption preferences does not draw `caption` or
    `subtitle` items, and a frame it draws without them is not the published frame.
- **`wai-vector`: the text layer.** `wai_vector::text` computes `text_hash`, block
  identities, effective windows, active sets, the projection, the transcript and the
  text record, in a sink's build, and a verifier's check of the cluster rule. The tool
  gains `wai-vector a11y` (the text record, the transcript, or one frame's projection)
  and `wai-vector text` (the text digests and each block as a reviewer reads it). The
  corpus-seeded fuzz holds the text layer of every payload it accepts: the transcript
  is JSON that gives back each tag and string, and a projection lists exactly the
  blocks the frame procedure draws at that tick.
- **`vector-conformance`: the `a11y` group.** Each payload's expected file states the
  SHA-256 of its text record, and the record is committed beside it, so every fabric leg
  recomputes it: the worked example of §5.4, windows and a fade to nothing, a
  `decorative` block, text alternatives of paths, reading groups, ruby over part of a
  base, registered and unregistered purposes, merged and clipped intervals, windows cut
  by enclosing groups, the three writing modes, four rates over one document, strings
  that take each escape and strings that take none, and a payload with no text. The
  `canonical` and `refusals` groups gain text-block cases: an ungrouped block between
  groups, ruby links to a later block and ranges that end inside a character or at the
  string's end, and pairs that pin where `TEXTS` sits in the record pass and that the
  text-block rules come before any outline set is read.

### Reader limits that bound memory: manifest values, and text decode

Two reader limits join the manifest length limit (SPEC §2, *Reader limits*), and with
them a reader's memory is bounded by what the specification states, not by what an
envelope asks for. `SECURITY.md` listed both as known gaps; it now states the bounds
and the tests that measure them.

- **Manifest values.** A reader need accept no manifest of more than 2²² JSON values,
  the manifest object and every value in it at any depth. The reference readers count
  the values before they build anything, refuse one past the limit
  (`manifest_too_many_values`), and build what they read at the size counted: at most
  128 bytes a value and one a byte on a 64-bit target, where a manifest of small
  values at the length limit took several gigabytes to parse. The stdlib Python
  verifiers refuse the same manifests, and the HTTP corpus holds a case at the limit
  and one past it.
- **Text decode.** A `wai.text.zstd` or `wai.text.xz` payload decodes to at most 2²⁷
  bytes and declares at most 2²⁷ bytes of window or dictionary; a reader may refuse
  more (`text_decoded_too_long`, `text_window_too_large`). The reference decoders
  write into a buffer never sized past the limit, refuse a Zstandard frame that
  declares a longer content size before decoding anything, and give the decoding
  library the window limit, so a declaration past it is refused from the stream's
  header. The reference encoders refuse input past the length limit. The C entry
  points `wai_text_zstd_decode` and `wai_text_xz_decode` are bounded the same way; the
  header gains `TEXT_DECODED_MAX_LEN` and `TEXT_WINDOW_MAX_LEN`.
- **Planned break for writers.** A conforming encoder writes no manifest past the value
  limit and no text payload past either text limit. No committed envelope or vector is
  past any of them, and no digest changes.
- **The codec wrappers in CI.** `wai-conformance.yml` gains `codec-wrappers`, on the
  macOS runner: the wrappers' fuzz smoke, the text decode limits and the C entry
  points, built with `codecs`. The step fails if any of its test binaries ran no test,
  which is what a build without the feature does.

### Integer operators: the reference crate, the bounds pass and its certificates

The crate, the tool and the corpus that
[integer-operators](extensions/integer-operators.md) (Draft) names exist. No capability
decodes through them yet: each capability's reference decoder is still the one its
payload extension names.

- **`wai-ops`**, the reference crate, `no_std` with `alloc` and no dependencies: the
  512-bit saturating integer of §8.1, the canonical JSON writer of §8.5, the `WHG1`
  parser and writer, the load-time and run-time plans, every graph operator of §5, the
  stage operators E1–E13, and the lowerings of §13.
- **Load step 8.** `wai_ops::bounds` runs the bounds pass of §8.1–§8.5 and writes the
  canonical certificate. The interpreter runs only a loaded graph, so nothing decodes a
  parameter set its certificate does not admit. `interp::run_observed` reports every
  value a run forms; `bounds::Within` holds a run to its certificate, and every golden
  decode runs that way.
- **`wai-verify`**: `bounds`, `lower` and `vectors`, with `--check`, `--witness` and
  `--advise`.
- **`wai/ops-conformance/`**: per-operator, stage, graph, load and lowering vectors, the
  lowered graph of every committed parameter set, and each set's certificate.
- **Two certifiers.** `tools/wai_ops_check.py`, written from the extension text and
  sharing no code with the crate, recomputes every vector and every certificate byte
  for byte.
- **`wai_ops::limb32`**, a reference model of the 32-bit arithmetic the `acc32` profile's
  two products need, checked against 64-bit arithmetic.
- **Model checking.** The operator set's harnesses are in `wai-ops/src/proofs/`, hosted
  by `wai-verify` and run by `wai-formal-verification.yml`. A harness that does not yet
  verify within the job's budget is kept, behind `cfg(kani_open)`, and is not run or
  counted as proved.
- **`wai-ops.yml`** runs the crate's tests and `vectors verify` natively, under
  wasm32-wasip2 and under wasm32 in Node, and runs the checker.

### wai-quantum 0.4.3: metered runs, placement by run, and the fabric wrapper

Additive. The default features, `full`, both implementations of the WebGPU member, and
every existing pin and digest are unchanged; runs, receipts and members stay unmetered
unless a caller supplies a meter.

- **Run meters.** `quantum_fabric::RunMeter`: `start` just before a run's first
  submission, `stop` after its last readback, returning the microjoules in the window and
  their `EnergyClass`, or none when the window cannot be attributed to the run alone. The
  trait is member-neutral, and no implementation ships in the crate. A reading reaches a
  result only when `check_label` accepts it; otherwise the run is unmetered (zero, no
  class). Every start is matched by one stop, a refused run's reading is discarded, and a
  meter that cannot start is never stopped.
- **The WebGPU member, metered.** `WebGpuMember::simulate_metered(c, start, meter)`
  returns the state with the accepted reading. Its window runs from the first submission
  to the last readback: the device's work, and the host's waits and copies between them;
  planning, buffer creation, the upload's staging and the conversion to a state vector
  are outside it, and the member's `detail` says so. A circuit refused
  before any submission opens no window. `member_metered(corpus, meter)` certifies as
  `member` does and gives each entry the figure its runs earn;
  `blocking::{simulate_metered, member_metered}`.
- **Placement figures from metered calibrations only.** `EnergyCalibration` (`new`,
  `add`, `runs`, `finish`) turns metered runs into a `KernelEntry::energy` figure: the
  total energy over the operations performed (`pair_updates`: one pair of amplitudes under
  one gate), in nanojoules rounded up, with the runs' class, whose uncertainty is the
  largest of the runs' plus what the rounding adds, over their windows together. An
  unmetered run, a refused label, a second class, or an uncertainty that no longer fits
  its field once the rounding is added leaves no figure.
  `detect_local_metered(corpus, meter)` gives the CPU reference its figures. Because
  `place` ranks a known figure before none, figures go on the CPU reference and on the
  members placement compares it with, or on none. `place` is unchanged; `place_for` and
  `Fabric` hold the rule themselves, and read figures on only some candidates as figures
  on none.
- **Placement for a run.** `place_for(members, kernel, required, predict_pj)` places by
  the least predicted picojoules when every candidate has a prediction, ties falling as
  in `place`. Otherwise predictions are not used, and the order is `place`'s, except that
  figures on only some candidates count as figures on none. Under synthetic models with a
  fixed cost per GPU run, a 3-qubit circuit stays on the CPU and a 24-qubit one goes to
  the GPU.
- **The fabric wrapper.** `Fabric` holds members with their `Runner`s (`CpuReference`,
  and `WebGpuMember` for its own implementation) and optional meters (`add`,
  `add_metered`, `set_model`, `order`, `run`). `run` places the circuit, falls back down
  the placement order when a member refuses with `MemberError::{TooWide, OutOfRange,
  Memory, Timeout, Device}` (`falls_back`), stops at an invalid circuit or a defect, and
  seals the `FabricReceipt` with the member and implementation that produced the result
  and its accepted reading. The refusals (`Refused`) are listed beside the receipt in
  `FabricRun`, or in `FabricError`; the receipt's signed bytes are unchanged.
  `WebGpuError` converts into `MemberError`. A runner states the member it is
  (`Runner::member_id`), and a run that reaches a member paired with a runner of another
  id is refused there as a defect, so a receipt cannot name a member that did not produce
  the result. `ready(future)` gives the output of a future that is ready on its first
  poll, which is how a native caller drives `run` without an executor; `&R` is a `Runner`
  and `&mut M` a `RunMeter`.
- **Tests.** CPU-side tests, native and under wasm32-wasip2 (each required by name in the
  `quantum-wasip2` job), cover the readings, the calibration, figures on both members or
  neither, `place_for`, the fallback, and metered fabric receipts. The device test
  `g13_metered_runs_and_the_fabric` runs the member under a test meter, calibrates it on
  random-v1 with the pinned digests, and runs the fabric with the member held to three
  qubits: a wider circuit falls back to the CPU reference, a narrower one is produced on
  the member and metered. It runs the tile implementation the same way, whose receipt
  names it, and requires a member paired with another member's runner, or with the other
  implementation's, to be refused. The member script requires it, and a fabric line per
  declared adapter.

### wai-quantum 0.4.2: tile runs, and the WebGPU member in a browser

Additive. The default features, `full`, the per-gate implementation and every existing pin
and digest are unchanged.

- **Tile runs: a second implementation, `wgsl-lanes32-tile/54e443520c020013`.**
  `plan::Config` (`PER_GATE`, `TILED`; `tile_size`, `run_cap`), `plan::Step::Tile`,
  `compile_with`, `schedule`, `execute_from`, `PlannedGate::{new, is_identity,
  is_well_formed}` and `TILE_BITS_MAX`; `PLAN_REVISION` is 2. A tiled plan applies each
  maximal run of consecutive gates that act inside a tile (diagonal gates on any target,
  others with their target inside the tile) to tiles of up to 1024 amplitudes, in
  op-log order, and leaves out exact identities only (circuit/1's `P(19..=32)` is
  `2^30 − 1`, not the identity, and stays). The CPU executor mirrors a tile kernel's
  index arithmetic and equals the per-gate program on any words; every configuration
  certifies with the lanes32 pins. A program whose gate's `diag` disagrees with its
  words, or whose op index is `u32::MAX`, is refused by the executor and by the member.
  `wgsl::TILE_SHADER` (`lanes32.wgsl` followed by `lanes32_tile.wgsl`, entry point
  `apply_tile`: one workgroup of 256 per tile, the tile updated in place in the state
  planes, a storage barrier between gates, the tile size in the run record, never in
  the text; a run record whose tiles or grid do not cover the state exactly once is
  refused) has its own id (`tile_implementation_id`), records (`tile_gate_record`,
  `tile_run_record`), grid (`tile_grid`) and mutants (`TILE_MUTANTS`, with the
  detectors `Race` and `Malformed`, `race_witness`, `malformed_tile_run`).
  `WebGpuMember::tiled()` gives the same device running it, probed and certified on its
  own; the member splits a run into dispatches of at most `SUBMIT_BUDGET` pair-updates.
  Discovery opens per-gate members, which are the fallback. Given both implementations'
  entries for one member and no energy figures, `place` breaks the tie by
  implementation name and places the tile implementation (`wgsl-lanes32-tile/…` sorts
  first). The checker holds `TILE_SHADER` to its own rules, barrier placement and
  barrier uniformity included, and its SPIR-V to the validator; the member's device
  tests, the member script and the browser leg require the four pinned digests under
  both implementations and every tile mutant caught. `tests/fabric_expected.rs` (not
  in the package) holds `quantum-conformance/fabric/expected.json` to the crate.
- **The member in a browser.** wai-web's feature `quantum_webgpu` binds the member for JS
  (`quantum_webgpu_discover`, `QuantumWebGpu`: `certify`, `statevector_hash_hex`) in a
  bundle of its own (`--out-name wai_quantum_webgpu`); the default build and its bundle
  are unchanged. `demo/quantum_webgpu.html` runs the member in a worker, its kernel
  compiled by the browser's own WGSL compiler. `quantum-conformance/browser_webgpu.mjs`
  drives the page headless and requires the four certificates bit-exact with the pinned
  digests (`quantum-conformance/fabric/expected.json`), every vector's statevector hash,
  every reject file refused with its class, and a circuit wider than the member holds
  refused. The `checks` job builds the bundle, and a `browser` job runs the page on macOS.

### wai-quantum 0.4.1: the fabric's WebGPU member

Additive. The default features, `full`, and every existing pin and digest are unchanged.

- **Certification one case at a time.** `quantum_fabric::Certification` (`new`, `add`,
  `cases`, `finish`) folds a member's outputs into a certificate as they arrive, for a
  member that produces them asynchronously. `certify` is now a loop over it and gives the
  same certificates.
- **Corpora.** `Corpus::Random { seed, cases }` is the existing random corpus
  (`random-v1`). `Corpus::Fabric { seed, cases, wide }` (`fabric-v1`) continues the same
  stream after it with circuits of 11 to 17 qubits (`lanes::random_wide_circuit`), so that
  every dispatch shape a 32-bit-lane member uses is exercised. `FABRIC_DEFAULT` is seed 7,
  200 + 24 circuits; `GENERATOR_FABRIC`, `GENERATOR_FABRIC2` and `fabric_generator` name
  it per gateset, and `Corpus::{generator, name, circuits}` describe any corpus. The
  fabric-v1 lanes32 digests are pinned: `41c726a0…3df7` (circuit/1) and `9a84b55a…7809`
  (circuit2). `lanes::certify_corpus` certifies the lane core on any corpus;
  `certify_lanes` is it on the random corpus, with the same generator string.
- **The lane core, word by word.** `lanes::GateWords` and `gate_words` (a gate as the
  eight 32-bit matrix words and the target and control masks a kernel reads),
  `fxmul_wide` (the exact 64-bit `fx`), `PROBE_EDGES`, `apply_words` and `run_words` (the
  general 2×2 path computed from exact 64-bit values, flagging a gate exactly when the
  reference's result leaves 32 bits, and refusing malformed words), `simulate_checked`,
  `output` and `reference_output`. `lanes::horizon(n, gs)` is the gate count below which a
  valid circuit provably cannot leave 32 bits, computed in integers from the gateset's
  table: 46,340 (circuit/1) and 46,339 (circuit2) at 26 qubits. The module documentation
  and `fxmul` no longer claim that every result of a valid circuit fits in 32 bits; it
  holds within the horizon.
- **`quantum_fabric::plan`**: `Program` (one step per op, the start reduced modulo 2^n),
  `PlannedGate`, `Step`, `compile`, `LaneRun` and `execute`, the CPU executor a member's
  output is held to; `PLAN_REVISION` is 1.
- **`quantum_fabric::wgsl`**: `SHADER`, the lane core transcribed into WGSL
  (`src/quantum_fabric/lanes32.wgsl`: unsigned 32-bit values only, an exact 64-bit copy
  of every result, literal shift amounts, no float, division, conversion, workgroup
  memory or barrier); `implementation_id()` (`wgsl-lanes32/dc8e19f61e0e87ab`, a BLAKE3
  over the shader text and `HOST_LAYOUT`, so any edit is a new implementation that earns
  its certificates again); `gate_record`, `Grid`, `grid` and `gate_grid` (the 64-byte gate
  record and the dispatch grid; `WORKGROUP` and `X_GROUPS_MAX` are 64); `AdapterIdentity`
  with `member_id` (`local/gpu-…` or `local/swgpu-…`, sorting after every
  `local/cpu-<arch>`), `key` and `class`; and `MUTANTS`, `mutate` and `Detector`, the
  shader mutants every device must catch. The module documentation carries the argument
  that the kernel is bit-exact. `wai/tools/quantum-wgsl-check` (unpublished) holds the
  shader to that argument's rules on naga's IR and translates it to MSL, SPIR-V and HLSL.
- **The WebGPU member, feature `quantum_fabric_webgpu`** (opt-in, not in `full`). The
  kernel through the platform's GPU interface natively (Metal, Vulkan, DX12) and through
  a browser's WebGPU on `wasm32-unknown-unknown`; wgpu 30.0.1 enters only those two target
  tables, and on every other target, `wasm32-wasip2` included, the module is compiled
  out. `discover()` opens each adapter at the WebGPU downlevel limits with only the buffer
  limits raised, pins the DX12 shader compiler (translator `naga+dxbc`), probes the
  device's arithmetic against the lane core, and lists each adapter it cannot use with
  the reason (`Discovery`, `Refusal`). `WebGpuMember` runs lanes, programs and circuits,
  certifies a gateset on a corpus, and builds a fabric `Member`; a bit-exact member
  reproduces the pinned random-v1 and fabric-v1 digests. A result that would leave 32
  bits is refused (`WebGpuError::OutOfRange`), never returned. Every entry point is async
  with no executor; natively `blocking` resolves each call in one poll, and a future still
  pending is an error. Natively each submission is waited on before the next, so
  `POLL_TIMEOUT` bounds one chunk of gates, not a run; a failed wait, which the GPU
  interface reports by panicking, is caught and returned as `WebGpuError::Device`, and
  the member treats its device as lost from then on (a `panic = "abort"` build still
  aborts). With no energy figures, placement keeps the CPU reference, and a receipt over
  a member's run is unmetered.
- **wai-rs.** The feature `quantum_fabric_webgpu` forwards to it and re-exports the
  module. `wai_quantum_vectors verify <dir> --webgpu` runs the corpus check once per
  WebGPU member, each state computed by the member; each accept file must also run on the
  member as it does natively, so the 255-qubit files are refused as `too-many-qubits`.
  Without the feature `--webgpu` is an error; with no member it exits 2.
- **CI.** `wai-quantum-webgpu.yml` checks the shader under naga and a SPIR-V validator,
  the browser build, the feature's absence from the `wasm32-wasip2` dependency tree, the
  shader file in the package, and the MSRV; and runs the member's tests
  (`scripts/wai-quantum-webgpu-member.sh`) on software Vulkan, Metal, and a discrete GPU
  through Vulkan. Each member job declares the adapters its runner must have, and fails
  if the adapters found differ, if any test did not run, if a declared adapter lacks a
  bit-exact certificate for each pinned digest, or if a shader mutant survives. The
  `wasm32-wasip2` job requires the fabric's new unit tests by name.

### Packaging: refinement layers, stills, a piece's ladder, live windows and delivery

- **Refinement layers in packages** ([packaging](extensions/packaging.md) §5.4). For
  each refinement rung of a piece's ladder the packer writes a `wai.video.int_refine`
  object over the piece's `wai.video.int_motion` base: the manifest `wai stage refine`
  writes, and a `WIR1` from `wai-rs`'s refinement encoder (`staged::wir1::encode`), one
  unit per frame, toward the source from the stage below as a sink reconstructs it. Each
  layer is applied to the decoded base by the reference sink before it is kept; the
  index states each prefix's tier, digest, SSE and deadline.
- **A piece's ladder, and stills** (§5.7). A rule's `then.ladder` chooses the rungs a
  piece is coded in; otherwise the component's first base rung, its refinement chain and
  its floors. §3.4 item 10 now requires a rule's ladder to name exactly one base, every
  refinement over a rung it names, and every floor of its component. A base of
  `wai.image.png` is a still: the piece's first frame, held for the piece, as a PNG every
  byte of which is stated (the `Sub` filter, one fixed-Huffman deflate block whose
  matches are stated), so an independent implementation writes the same file.
- **Catalogs** (§8.3, §8.4). The MoQ catalog is now a stage set of moq-streaming-format
  §4.1, written by `wai::staged::moq`: `wai_staged` on every track, a receipts track and
  an alias per media track, `depends` on a track of every layer below; it replaces
  `priority` (MOQT's own, §8.5) and `wai_layer`, which the registry no longer lists. Over
  HTTP, whose variants cannot mix bases with no capability in common (http-delivery §4.2),
  a still's piece is played from the floor; its stage view and MoQ tracks list it.
- **Live windows** (§9.2). `wai-pack live` packs a Y4M stream from standard input as it
  arrives, with the coder the on-demand packer uses, in windows sealed by version-2
  receipts chained `live-window`; its bases and refinements are the on-demand package's,
  byte for byte. `wai-pack listing --primary` lists what the two share. It writes as it
  goes: each object when it is published and each window when it is sealed, every file
  whole, and the plan file and `LISTING` at the end. It does not keep an object's bytes
  once written, or a record per file: each listing line goes to a hidden file and is
  sorted into `LISTING` at the end. A source that fails mid-stream keeps every sealed window and the objects
  it lists, removes the rest, and exits 1 saying what it kept.
  `wai::pin::listing_from_sha256` writes a listing from per-file digests.
- **Stills read by PNG's rules** (§5.7). Both reference readers take a PNG of 8-bit RGB
  without interlace and agree on what that admits. They read a suggested palette,
  ancillary chunks and split image data. They refuse an interlace, bytes after `IEND`
  or after the zlib stream, a short row, a zlib window past 32 KiB, and every other
  chunk fault (`stills/`). Both inflate as the zlib library does: no over-subscribed
  Huffman code, and no incomplete one but a lone one-bit code. wai-pack's reader has
  its own inflater, and `png_diff.py` runs it against `pack_verify.py`'s zlib.
- **The longest piece** (§3.4, §4.4). `max_ms` is at most ten minutes, and at most
  `2^24` units on every component's timebase (`policy-piece-bounds`). A verifier
  refuses a piece longer than the policy's `max_ms` before it decodes a frame. It
  checks each floor object's unit count before decoding (§11.3). It hashes a still's
  frame held for its piece without repeating the frame.
- **Manifests checked by both verifiers** (§11.3). A base's manifest must be §5.2's and a
  refinement's §5.4's (`object-manifest`, a new reason). A base that does not decode
  is `prefix-digest`.
- **A repeated object over HTTP** ([http-delivery](extensions/http-delivery.md) §4.2).
  A repeated object contributes once, with its layers, and a layer joins every
  occurrence of its base. `http_verify.py` and `wai-rs` now agree on a playlist's
  codecs and floor where a piece repeats. `pack_verify.py` checks every golden playlist
  with `http_verify.py`'s union, floor and adjacency checks.
- **One stall bound.** `stage_measure::sim::stall_bound_for` is the only bound: a
  model's, which takes each refinement's own timeout. wai-web's simulator prints it, as
  the native tool does.
- **Delivery, simulated** (§8.9). `wai_pack::delivery` and `wai-pack sim` run
  staged-measurement's simulator over what a live packer publishes and what a class
  subscribes to: per-track delivery timeouts, §8.5's priorities, the subscriber's cap,
  group receipts on each track's chain. The simulator's model gains a refinement's own
  `timeout_us`, and §10.4's bound takes the longest refinement timeout
  ([staged-measurement](extensions/staged-measurement.md) Appendix B); its corpus adds a
  per-track model.
- **The corpus**: the golden `video-staged-01` (a moving shot, a bumper played twice, a
  still shot, a second moving shot; two refinement layers, a still, the event's deadline,
  three classes), its supplied floor and how it was written (`supplied_floor.py`),
  `staged-mutations/`, `delivery/`, `stills/` and `live/` (the golden packed live in
  windows of two pieces, whose chain `pack_verify.py` checks). `stage_diff.py` runs
  refinement layers, stills (wide ones among them, so every deflate distance code is
  used), piece plans and deliveries differentially against `pack_verify.py`, which
  re-derives every layer and still from the source byte for byte. `sim_diff.py` runs the
  delivery simulator differentially against `sim_verify.py`. `pack-refinement-unsupported` and
  `pack-rule-ladder-unsupported` are gone.
- The base golden `video-base-01` keeps its objects, index and receipt; its MoQ catalog
  and its `LISTING` change.

### The record made true: training, paths, protected content, reader limits, security

- **`int_synth`'s weights are quantised after float training.** SPEC §5 and
  [determinism-tiers](extensions/determinism-tiers.md) §5 said the capability decodes
  "QAT weights". No committed model was trained that way: the trained corpus vector,
  `photo`, was made by `tools/wai_int_synth_real.py`, float training then post-training
  quantisation with each layer's shift derived from its measured activation range, and
  the weights of the synthetic vectors and of the cross-architecture golden come from a
  seed. Both texts now say so, and so do the tools:
  `tools/wai_int_synth_qat.py` is marked experimental (it trains against one fixed shift
  and made no committed model), and `tools/wai_int_synth_export.py` names the working
  path.
- **Stale paths.** The int_synth decoder is `int_synth.rs` at the crate root, not
  `neural/int_synth.rs`. The quantum-ops engine and the quantum-sim receipt live in
  `wai-quantum`, whose features `wai-rs` forwards. README and SPEC Appendix A no longer
  send macOS builders to a `.cargo/config.toml` that no longer exists: `wai-rs/build.rs`
  finds libFLAC.
- **Protected content: two lanes** (SPEC §6.5). Lane A, hardware-protected content,
  decodes inside the platform's protected media path; WAI's records about it (among
  them the served manifest, the segment map, the key-release decisions, the ad
  decisions and impressions of `wai.video.ssai`, the program-as-run windows of
  `wai.video.channel`, and the content hashes of WAI objects) cover no decoded frame.
  This revision registers no `security.encryption` value for it. Lane B, end-to-end
  encryption at the application layer (`mls-sframe`), decrypts in the application, so
  the tier and the output digests of the sink that holds the key hold as they do for
  plaintext; staged-measurement counts no encrypted object, and a verifier without the
  key checks none of a `reconstruct-session` claim's rules that need a payload. Lane B
  is not a protected-path robustness level. `wai.video.keys` and the confidentiality
  extension say which lane they belong to.
- **Where the exact tier holds** (determinism-tiers §5). `decode-equivalence` is shown
  on native CPU and on wasm32, on no GPU or NPU; a WebAssembly decoder at the tier uses
  no relaxed-SIMD instruction and depends on no NaN's bits; and the tier never covers
  output a sink cannot read.
- **Reader limits** (SPEC §2). A reader need accept no manifest longer than 2²⁷ bytes
  (128 MiB) or nested deeper than 127, and refuses a longer one from `man_len` alone
  (`manifest_too_large`). The length limit is sized to the largest manifest a
  conforming staged layer has: a `prev` (staged-delivery §3) naming a different
  producer for each of the 2²⁰ units a `WIR1` layer may cover (staged-delivery
  Appendix A.3) is about 94 MB. It bounds what a reader reads, not the memory a parse
  takes, which in the reference readers is up to nearly 40 times the manifest's length
  (`SECURITY.md`). `wai-rs`'s readers, its HTTP head parser (`head_malformed`) and the
  Python verifiers apply the length limit, and the HTTP corpus holds it at its edge;
  the writers, `wai-web`'s included, refuse to write past it, and refuse a payload whose
  length does not fit its `u32` field.
  *Envelopes this makes non-conforming:* none of the layout's; an envelope whose manifest
  is longer than 128 MiB, or nested deeper than 127, is one a conforming sink may refuse,
  so a conforming encoder no longer writes it (SPEC §7).
- **`SECURITY.md`.** How to report a vulnerability, and where the reference
  implementation reads untrusted input: the manifest limits and the memory a manifest's
  parse takes, the native codec libraries and their gaps (unbounded decompression output
  for `wai.text.zstd` and `wai.text.xz`; no fuzzing in CI), module authorisation, and the
  fingerprinting risk of `held_sets`. CI checks its section references.

### Vector graphics (Draft): the extension text of `wai.graphics.vector`, §1–§3

- **A new extension, [graphics-vector](extensions/graphics-vector.md) (Draft).**
  `wai.graphics.vector` carries titles, credits and other authored graphics as integer
  drawing instructions, which the sink rasterises into an overlay layer. This revision
  writes §1–§3: overlay-equivalence and text-equivalence, which cover the overlay layer
  and the computed text and nothing after them; the `WVG1` payload and the `WOS1`
  outline-set format, with their canonical rules, caps and the order in which a sink
  refuses, including a per-frame cost each listed size states and a sink verifies at
  load; the digest sidecar; and the arithmetic, as the raster stage family RS1–RS9, in
  which no value wraps. Outline sets, text and accessibility, and packages follow in
  later revisions of the Draft. Nothing is registered: the string is registered once its
  conformance corpus is complete.

### wai-web on older engines: the reach builds

- **Three builds from the shipped package, and a loader** ([wai-web/reach](wai-web/reach/README.md)).
  `reach/build.sh` derives them from `wai_web_bg.wasm` and `wai_web.js` with no change to
  the Rust:
  - **L1** legalises the binary's JS interface (`wasm-opt --legalize-js-interface`), so
    engines where i64 does not cross as `BigInt` run the full module;
  - **L2** lowers it to MVP WebAssembly and legalises it;
  - **J** translates the MVP module to ES5 JavaScript with `wasm2js`.

  L2 and J come in profiles, kept by `wasm-metadce` from the exports each needs: `full`,
  `verify_decode`, `verify` and `receipt`. An ES5 runtime supplies the missing builtins
  and joins and splits i64 values at the boundary: a `BigInt` where the engine has one, a
  `Number` (exact below 2^53) where it does not. The glue is the shipped glue compiled to
  ES5, with counted changes. One of them rebuilds a memory view whenever the memory's
  buffer changes, not only when it is detached.
- **`wai-load.js` chooses by feature probes alone.** Tiny modules passed to
  `WebAssembly.validate` (MVP, sign-extension, non-trapping float-to-int, bulk memory), one
  instantiated module that returns its i64 argument, and checks for typed arrays decide
  between shipped, L1, L2 and J. The probes never throw. An early WebAssembly engine line
  (WebAssembly and the stack-trace capture API, but no RegExp `dotAll` flag: the feature
  `noDotAll`) gets J, or the L2 receipt build, because the full MVP module has crashed an
  engine of that line on arm64. A rung can be forced, and the loader says why it chose.
- **Pinned and recorded.** binaryen `version_123` is checked against each platform
  archive's SHA-256, and TypeScript, esbuild, acorn and binaryen's WebAssembly build are
  pinned by `package-lock.json`; J's `wasm2js` is that WebAssembly build, so J comes out
  the same on every host.
  `reach/manifest.json` records the SHA-256 and sizes of every output built from the
  package the site serves, and the commit that package was built from; `reach/record/`
  keeps that package. Given it, the build refuses any output that differs.
- **Checked output for output.**
  - `reach/test/conformance.js` (ES5) runs workloads over the repository's corpora on any
    rung, and every output must equal the shipped build's. The workloads cover receipts,
    the staged corpus, the integer payloads, int_synth, the menu, the output tree and the
    i64 boundary.
  - `reach/test/run-modern.mjs` forces every rung and profile on a current Node, and runs
    the demo verify scripts against each full build.
  - CI (`wai-web-reach.yml`) runs a leg per package. On the package of record, kept in
    `reach/record/` and held to the manifest's SHA-256, every output must equal the
    manifest's; on the tree's own wai-web, the reach builds must still build and pass.
    Each leg runs the conformance on the Node images of the engine lines each rung serves
    (J on `node:0.12`, the early-engine routing on `node:8.0.0`, L2 on `node:8` and
    `node:10`, L1 on `node:12` and `node:14`) and the shipped build on a current Node.

  The images stand in for device classes, which the README names by capability. No real
  old device has been run.
- **The receipt vector's signer id is a label.** The receipt workloads verify an asset and
  its provenance receipt in `reach/test/vectors/provenance/`, signed with `wai_prov`'s
  published test seed under its label `did:example:wai-prov`. `wai_prov --carry
  <receipt.json>` measures nothing: it carries each step's figure, and its label, from an
  earlier receipt for the same lineage and signs them under this revision's id, and it
  refuses a receipt that does not verify or whose steps are not this run's. The vector
  carries the figures `wai_prov` measured when it first wrote the asset, and the
  documented command, carrying the receipt onto itself, writes both files again byte for
  byte.

### Generated pictures and scores, made natively and checked by SHA-256

- **The payload of `wai.generate.spectral`** (SPEC §5, *Payload of
  `wai.generate.spectral`*). The registration is completed (§8): the payload is an
  instruction object of form v0, one canonical JSON object naming the model by its
  SHA-256 and giving the prompt, the seed, the steps and every size a sink makes. One
  serialization per object, so an envelope's content hash names the instructions; a
  sink makes only listed sizes, with that model only, and checks each output by its
  SHA-256, never by a 64-bit FNV-1a alone. The intent of an envelope that names the
  generator is `create`, and an envelope with another intent is non-conforming. The
  model is the generator's parameter set (§3.1) and `model` its identity, so an
  envelope that names the generator carries no `prior` pin, and a sink refuses one that
  does (§3.1 says so beside the pin rules). `wai inspect` and `wai unwrap` report the
  payload, the intent and any pin of such an envelope, ✓ or ✗ with the reason. Vector:
  `wai-rs/tests/vectors/generate/instr-v0.json`, the object, the SHA-256 at each of
  its four listed sizes, twelve objects at the inclusive bounds a sink accepts (each
  with the SHA-256 of its first listed size), and 36 objects a sink refuses, each with
  the rule the reference reader finds broken first, which its tests assert.
- **`wai generate`** (feature `generate`) makes a picture with the reference generator
  at any size form v0 allows and prints its SHA-256; `--instr` makes a listed size of
  an object, read as a sink reads it, and `--write-instr` writes one. `sizes.json` pins
  the committed prompt at 256×128 and 640×320. The wasm32 workflow checks that the
  browser binding makes every pinned picture and score with the SHA-256 the native
  tool wrote (`wai-web/demo/verify_generate_wasm.mjs`), so a page can hold its own
  output to these pins. `wai wrap` refuses a `wai.generate.spectral` payload that is not a
  form-v0 object, and any intent but `create` over that generator, in every build.
- **`wai score build|render`** (feature `score`). `build` writes a `WSCR` from integer
  note fields only (score §2.3); `render` prints `samples`, `pcm_sha256`,
  `mixdown_blake3` and `score_hash`, and `--pcm` writes the samples. Every
  `score-conformance` vector now states `pcm_sha256`, and `notes/<name>.json` holds it
  in the authoring form, which `build` turns back into the committed `WSCR` byte for
  byte.
- No library source changed: the generator, the score synth, the world floor and the
  browser binding are as they were, so pins made with this tool bind the binding pages
  already load.

### Quantum circuits: `wai.quantum.circuit2`, and `wai.quantum.circuit` stated exactly

- **A new capability, `wai.quantum.circuit2`** ([quantum-sim](extensions/quantum-sim.md),
  SPEC §5). Magic `WQC2`, contract `ext` `wai.quantum.circuit2/1`, numeric
  `wai.det.amp64`, gateset `cliffordT+dyadicP.rn`. The container layout, the gate IR and
  the arithmetic are those of `wai.quantum.circuit`; the phase table is not. Each
  `P(k)` holds the integers nearest to `2^30·cos(2π/2^k)` and `2^30·sin(2π/2^k)`, so its
  controlled-phase ladder gives the QFT. The literal table is normative (§3.2, Appendix
  B.2, digest `d34cb793…3510`); an exact-integer interval procedure certifies every entry,
  informatively. The contract and the measurement are byte-canonical, and the op-log ends
  exactly after its last op. Registered in the machine-readable registry and the
  vocabularies, with `wai.det.amp64` as an identifier.
- **`wai.quantum.circuit` is frozen, and its table stated for the first time** (§3.1,
  Appendix B.1, digest `3fa9805f…8466`). The text gives the exact recurrence — the sine
  formula, the truncating halving, the floor square root — and the deviation of every
  entry from the ideal phase: from `P(8)` on the imaginary part is too large, by up to
  7032 units of `2^−30`, and `P(19..32)` apply no phase. The claim that this table gives the
  QFT is withdrawn. Writers SHOULD NOT emit new circuit/1 content.
- **Both capabilities are specified to the bit** (§2–§4): little-endian integers, section
  bounds computed without overflow, a repeated known section refused, the contract and
  the measurement read as I-JSON (RFC 7493) by rules stated in full: no repeated member
  name, no surrogate or noncharacter, no number of magnitude 2^1024 − 2^970 or more, at
  most 127 levels deep, and an integer read exactly; the three refusal
  classes (malformed, contract, invalid) in the order a reader checks them, the
  complex-product composition, the index convention, the gate matrices, the hash layouts
  and the shot sampler. `circuit_hash` covers the canonical contract, so it names the
  capability and its table; `statevector_hash` does not. §5 states the error per
  capability, and a new §8 why circuit2 is a capability of its own and what an older
  reader does with it (it refuses `WQC2` at the magic).
- **Newly non-conforming circuit/1 payloads** (quantum-sim §8, SPEC §5): a contract whose
  `ext`, `numeric`, `frac` or `gateset` is missing or carries another value; an
  `n_qubits` that is not an unsigned integer of at most 255; a known section kind more
  than once; section bounds that overflow; a contract or measurement that repeats a
  member name within one object, holds a noncharacter, or holds a number of magnitude
  2^1024 − 2^970 or more that an earlier build's float parser let through (some builds
  also refused numbers just below it, which are now accepted). No file a reference writer
  wrote is among them.
- **Corpus.** `quantum-conformance/` keeps its four original vectors and manifest entries
  byte for byte and adds qft5_x22, dyadic_ladder and qft8_ones, which see the phase table:
  under either table, `dyadic_ladder`'s statevector hash changes with every ±1 change of
  any `P(1..32)` component. `v2/` holds the same seven circuits as circuit2; `reject/` at
  least one payload for each refusal a payload can carry, with its class; and `accept/`
  payloads at the edges of the same rules that a reader must accept, with how each reads.
  `tools/referee.py`, integers only, derives both tables, requires Appendix B to state
  them, reads every JSON section by the same rules, and rebuilds and checks every file. The
  Appendix B tables are rendered from the reference constants by
  `wai_quantum_vectors appendix`, and a test fails when the text differs.

### wai-quantum 0.4.0: circuits carry their gateset

- **`Gateset { V1, V2 }`** names the capability a circuit is reconstructed with, and
  `Circuit` carries one in a new field. `Circuit::new` and `Circuit::qft` build V2
  (`wai.quantum.circuit2`); a circuit read from a container takes the container's, and one
  derived from another keeps it. `PHASE_V1` and `PHASE_V2` are literal constants: no
  recurrence runs at load. `simulate`, `to_wqc`, `circuit_hash`, receipts, the fabric's
  lanes and the toolchain follow the circuit's gateset. `from_wqc` reads both magics,
  checks a circuit/1 contract's values and a circuit2 contract's bytes, reads the contract
  and the measurement as I-JSON with a strict reader of its own that decides every rule
  from the text, the number range on the literal's digits (so neither a build feature nor
  a float parser's limits changes a verdict), refuses repeated sections and
  wrapping bounds, and no longer sizes an allocation from a claimed op count.
  `reject_corpus()` and `accept_corpus()` return the payloads a reader must refuse, each
  with its class, and must accept, each with its reading; the corpus's `reject/` and
  `accept/` trees are written from them.
- **Every circuit built in code changes its `circuit_hash`**, since the contract names the
  capability, and its statevector wherever `H`, `T`, `T†`, `P(3)` or `P(k ≥ 5)` acts on a
  nonzero amplitude (`P(1)`, `P(2)` and `P(4)` are the same in both tables). Every
  circuit/1 artefact reproduces: the four original vectors, the legacy receipt pins, the
  lanes32 certificate corpus and the other pins are unchanged, built in `Gateset::V1`, and
  new V2 pins sit beside them.
- **wai-rs** registers `wai.quantum.circuit2`; `quantum_envelope` tags an envelope with the
  circuit's capability and refuses one whose capability and magic disagree. The
  `wai-quantum` CLI takes `--gateset v1|v2` (default v2) and names the capability in its
  JSON; with `--gateset v1` and a phase of `k ≥ 8` it prints the circuit/1 deviation on
  standard error. `wai_quantum_meter` seals the capability the file names. The browser
  sink reads both magics; the IDE has `set_gateset`, and its `add()` sets the parameter of
  every gate but `P` to 0. The serve component, `wai-quantum-serve` 0.5.0, takes
  `gateset=v1|v2` on `/run`, `/qasm` and `/receipt`.
- **Migration from 0.3** (a breaking release):
  - `Circuit::new` is V2. To reproduce or verify a 0.3 artefact, read its `WQC1` file, or
    set `c.gateset = Gateset::V1` on a circuit built in code; a receipt's `gateset_of`
    says which gateset a circuit's ops match it under.
  - `quantum::phase`, `quantum::inv_sqrt2` and `quantum::CAPABILITY` are removed in favour
    of `Gateset::phase`, `Gateset::inv_sqrt2` and `Gateset::capability`.
  - `to_wqc`, `circuit_hash` and `quantum_toolchain::circuit_id` return a `Result`; they
    fail only on a circuit its gateset does not allow, never on one that has simulated.
  - `Circuit` and `QuantumError` are `#[non_exhaustive]`; `QuantumError` gains `Contract`
    and `kind()`.
  - `quantum_fabric::certify_lanes` and `lanes::random_circuit` take a gateset, with a
    kernel and generator id per gateset; the circuit/1 ids are unchanged.
  - `validate` refuses a gate with more than 255 controls, and in V2 repeated controls and
    a parameter on any gate but `P`. `from_wqc` refuses the payloads quantum-sim §8 lists.

### Integer operators (Draft), and the payload text of two integer codecs

- **A new extension, [integer-operators](extensions/integer-operators.md) (Draft).**
  `wai-ops/1` is the integer operator set the integer-exact capabilities decode
  through: each operator's exact arithmetic over the integers, with rounding, bias
  placement and shift scope as parameters of requantisation; a range policy on every
  tensor (`derived`, `clamp` or `check`); the `acc64` and `acc32` accumulator profiles;
  static bounds rules, saturating at `±2^255` so that a 512-bit integer holds every
  value they form, and a canonical bounds certificate that decides which profile a
  parameter set meets; the accelerator subset `wai-ops/1-npu`; the `WHG1` graph form;
  the stage operators of the drivers, among them the mANS decode of
  `wai.image.jpegai` (E8) with its scale path (§13.6.1); and canonical lowerings of
  the existing network forms (`int_synth.json`, `WHM1`, `WAM1`, `WMB1`) to `WHG1`.
  The reference crate, the tool and the conformance corpus it names follow this
  Draft; until they land, each capability's reference is the decoder its payload
  extension names.
- **[integer-payloads](extensions/integer-payloads.md) writes down two more formats.**
  §13 is the `wai.neural.int_synth` payload and its set, `int_synth.json`; §14 is the
  `WIM1` payload and the `WMB1` bundle of `wai.neural.int_mlicpp`, including that the
  entropy network reads every element of `y` not yet decoded as 0, and that the three
  networks and the tables share one `afrac` in `1..=30`. §9 gains the members and
  checks of `entropy.json`, and states that the layers' arithmetic is
  integer-operators §13.2 and §13.3. No decoder changes: §13 lists the
  `wai.neural.int_synth` checks the reference decoder makes once it decodes through
  its lowering.

### Staged measurement and the session claim in the browser

- **The measurement core, bound for JS** ([staged-measurement](extensions/staged-measurement.md)).
  `wai-web` binds `wai::stage_measure` (`src/measure.rs`), the code `wai_stage_measure`
  runs, compiled to wasm32: `evaluateTitle` gives the text `eval --json` prints for a
  title (counted bytes, quality integers, first presentable bytes, the layering
  overhead's points, deduplication, the output tree), `simulate` and `simulateModel` the
  text `sim` prints, `countObject` one envelope's counted bytes or the code it is not
  countable with, `qualityRgb8` one picture pair's integers, and `outputRoot` the output
  tree's root. A title is read from a map of files; `titleMissing` names the next file it
  needs, so a page fetches them one at a time. Sums of squares are decimal strings, as the
  corpus writes them.
- **The session claim, sealed in the browser** ([jwp-receipts](extensions/jwp-receipts.md)
  §2.4.12). The claim classes (`receipt_claims`) build for wasm32 with the measurement
  core, and CI builds them so. `sessionBegin` records a session at its entry object
  (`wai::session_recorder::SessionRecorder`): the decoder that decoded each envelope, each
  presentation with its output's digest, each object received. `finish` seals the claim
  with a key whose seed the page supplies, checked against the record first; a browser
  session's `energy` is `null`. `verifyClaim` checks a claim against the presentations and
  the objects as the claim-conformance runner does.
- **Proved byte for byte.** `demo/verify_measure.sh`, run in the wasm job, requires every
  golden, refusal title, counted object, quality pair and simulator result of
  `staged-measure-conformance/` from the wasm32 build, and the native `eval --json` of every
  fixture to equal the wasm32 evaluation. A session recorded the way a page records one is
  the native tool's claim byte for byte and commits to the golden output tree; the wasm32,
  native and Python verifiers give the same verdict on it and on its altered variants.
- **A recorder that holds the envelopes checks the claim against them.** A sink hands
  `SessionRecorder` each object's bytes (`received_envelope`): the recorder derives its
  content hash, counted and transferred bytes, and whether an entry names it, and
  `seal_checked` checks the claim against every envelope it holds, as a verifier holding
  the `envelopes` context does, so an entry naming a capability its envelope does not give
  never leaves the sink (`session_source_mismatch`). An object the sink does not hold is
  `received_unheld`, by its content hash; the recorder no longer takes an object's figures
  as told. The native tool's session over a title and the browser's `SessionHandle` both
  hold their envelopes; `decoderFor` refuses a capability that is neither the envelope's
  nor its fallback and requires the decoder's `first_unit`, `received` takes the bytes
  alone — bytes that are not an envelope included, recorded as an inert object neither
  counted nor presented, so the session accounts for every byte transferred —
  `verifyClaim` takes the envelopes, and integers from JS are refused unless whole and in
  range.
- **An object is listed once, as its envelope gives it** (§2.4.12, the objects). No
  (content hash, path) is listed twice (`session_object_duplicate`): a sink that fetches
  an object again lists it once, its transferred bytes every byte of every fetch, as the
  recorder merges repeats. After the sums, an object is `presented` exactly when a decoder
  entry names it (`session_presented_mismatch`). With the envelopes, an object listed with
  a held envelope's content hash states that envelope's counted bytes, `null` exactly when
  it is not countable (`session_counted_mismatch`).
- **Envelopes no entry names are received, not refused.** Inert bytes — of neither
  magic, or whose `security` member is not a security object — are refused as unreadable
  only where a decoder entry names them (`session_envelope_unreadable`); those, and an
  encrypted envelope, no entry names are a received object, uncounted and not presented,
  and one listed with counted bytes, even 0, is refused (`session_counted_mismatch`). A
  page that receives such bytes lists them and seals its session.
- **The layers and units an envelope gives.** A decoder entry states `first_unit`, the
  unit at which its object's first unit is presented. With the outputs and the
  envelopes, each presentation through an entry whose envelope is held is at the layer
  count that envelope gives — 1 for a base or a rendition with no `refines`,
  `refines.layer + 1` for a layer (`session_layers_mismatch`) — and of a unit it covers:
  from `first_unit` plus the span's start, the span's count of units, or the base's units
  where a layer has no span or the object is a base (`session_unit_outside_object`). A
  re-decode decodes unit `unit − first_unit` of the entry's object. A base's units are read
  in its payload's own format — of its manifest capability's and its fallback's pinned
  formats, the one whose magic the payload starts with — whichever capability the entry
  names. `claim_verify.py` reads
  the units from the integer payloads as `wai-rs` does, and `differential.py` adds random
  sessions holding a base's and its layers' envelopes, as `WAI1` envelopes and as `WAI2`
  components.
- **A layer is one a sink can apply, at its base's first unit** (§2.4.12). An entry naming
  a held layer a sink refuses is refused (`session_layer_not_applicable`): not a
  refinement or with a fallback, its `refines` malformed or absent, a `WAI2` refinement
  component that does not conform (SPEC §7.1), its base a layer, not admitted by the
  layer's capability or decoded only through its fallback, its span or runs outside a
  held base's units, or its payload not `WIR1` with a unit for each it covers
  (staged-delivery §4.1 steps 1–4 and 6). A `WAI2` component and its exploded envelope
  are one base in either form (staged-delivery §9). A held layer's entry states the
  `first_unit` of an entry it is applied over — naming its base through its own
  capability — and an entry naming a `stage` claim's base or a layer it lists states the
  unit `stage_sets` names that claim at (`session_first_unit_mismatch`).
- **An entry names what a sink dispatches** (§2.4.12 rule 4): never a `WAI2` companion,
  a rendition or refinement whose payload does not match its `sha256`, a capability that
  is not dispatchable, or the fallback of a `WAI1` whose capability is companion-only
  (`session_source_mismatch`).
- **One verdict for either name, by capability, and for a key holder** (§2.4.12). An entry
  naming a `WAI2` rendition by its exploded envelope's content hash, with only the `WAI2`
  held, is read as naming the rendition by rules 4–8; the exploded envelope's own bytes,
  held, are a separate delivery. A manifest or rendition is a layer only when its
  capability is refinement-only: a base carrying a stray `refines` is a base, as
  dispatch and selection already treat it (SPEC §4, staged-delivery §3), and `wai stage
  apply` now routes by capability. An encrypted envelope an entry names is accepted —
  a sink holding its key decrypts and dispatches it (SPEC §6.4 step 3) — and what the
  entry states of it is the signer's statement.
- **A WAI2 offers its refinements only when every component conforms** (SPEC §7.1,
  *Selection*). The text now states the reference sink's rule: an envelope one of whose
  components breaks a rule of §7.1 or §5 (`check_component_roles`, now on `ManifestV2`)
  offers none of its refinements, and may still present the rendition it picked. `wai
  stage explode` follows it, refusing to explode a component of such an envelope that is
  of a refinement-only capability or marked a refinement; rule 5 matches it
  for the whole envelope; and http-delivery §7 step 4 states it for a client planning
  from a verified head: each refinement of a nonconforming envelope is dropped
  (`components_nonconforming`), never requested, one stating no `sha256` dropped for that
  too, on the range path and the whole-object path alike. A refinement is one whose
  `refines.base`, read on its own, is the pick's id (`container::refinements_naming`), so
  one whose `refines` is otherwise malformed is reported too; of a conforming envelope
  that is exactly the refinements whose `refines` is well formed. The HTTP corpus adds
  ONC's conforming twin, OCF, whose refinement is fetched and delivered on both paths,
  and ONM and ONX, a `refines` of the pick's id alone (reported) and of no entry's (not
  reported); `OR`'s refinement, which names the pick, is now reported. `falsify/` adds
  `conforming-withheld` and `refines-read-whole`, each with its `http_verify.py
  --mutate`.
- **An encrypted envelope's clear parts are checked** (§2.4.12). SPEC §6.4 encrypts
  payloads only: rules 4–9 now read an encrypted envelope's manifest and rendition table
  as any held envelope's — the rendition an entry names, its capability, the component
  rules, a base's layer count, a companion-only capability, a fallback, and the classical
  length its table declares — and leave to the signer only what needs a payload: whether
  a payload matches its `sha256`, a base's units and a layer's `WIR1` payload. A `sha256`
  that is not 64 lowercase hex digits matches no payload, so an entry naming its rendition
  is refused, encrypted or not (rule 4, `session_source_mismatch`).
- **The recorder states the classical length.** `SessionRecorder` derives
  `classical_declared` at the tally from the `WAI2` envelopes it holds, readable or
  encrypted, exactly as rule 9 counts them (`claim_receipt::session::classical_length`,
  which rule 9 calls too); `SessionRecorder::classical`, which a sink called by hand and
  the browser binding never could, is removed. A page that receives a `WAI2` it does not
  decode now seals its session. `verify_measure` seals two such sessions in wasm32 — over
  `counting/wai2-classical.wai`, whose second rendition is PNG, and its encrypted twin,
  both new in the counting set — checks the length against the rendition table, and the
  native and Python verifiers give the same verdicts. A limitation of this revision: the
  length comes from the `WAI2` envelopes held in full, so a sink that received a `WAI2`
  only as byte ranges, holding its head but not the envelope, cannot yet state that
  envelope's classical rendition (§2.4.12, `classical_declared` and rule 9). Recording
  such a session is HTTP session work still to come.
- **A record or a derivation is no component in any role** (SPEC §7.1, *Component
  identity*; `record-or-derivation-as-component`).
- **A record or a derivation is no companion** (SPEC §7.1, rule 1). A companion's
  capability is companion-only or unregistered; a capability registered in another
  kind — a payload, a record or a derivation — is refused
  (`standalone-as-companion`), which the reference now reads from the registry rather
  than from its list of media capabilities, which left records and derivations out.
- **Every refusal accounted for.** The claim-conformance README's table lists each
  refusal a sink makes that decides whether a presentation can happen — staged-delivery
  §3, §4.1, §4.3–§4.5, §9; SPEC §4, §5, §7.1 — with the §2.4.12 rule and vector that
  refuse it, or the sentence of §2.4.12 that leaves it to the signer's statement.
- **The browser claim is a release build.** `verify_measure_wasm.mjs` requires the wasm32
  claim's `host.build.profile` to be `release` and its target `wasm32-unknown-unknown`.
- **The build profile is the optimisation level.** `release` is a build at `opt-level`
  2, 3, `s` or `z`, whatever its profile is named, and `debug` one at 0 or 1
  ([jwp-receipts](extensions/jwp-receipts.md) §2.4.8's `build`): a WebAssembly build
  optimised for size is `release`.

### One Ed25519 rule for every signature

- **SPEC §7 *Signatures*.** Every Ed25519 signature WAI defines is verified strictly: the
  key is its point's one canonical encoding and not of small order, `S` < L, `R` is not of
  small order, and the cofactorless equation holds with `R` compared byte for byte. The
  receipt extensions cite the rule instead of restating it; jwp-receipts §3.1 step 3
  keeps its two reasons, `bad_signer_key` for steps 1 and 2 and
  `signature_verify_failed` for the rest. OER/2's chain signature keeps the rule
  oer-receipts §5 settles.
- **One verifier in wai-rs.** `wai::sig::verify_strict` is the only Ed25519 check written
  in the crate but OER/2's. The video, SSAI, key-release, channel and package receipts
  (both versions), the provenance, ternary, spatial-audio, volumetric, film, score and
  haptics receipts, and a world's session receipt, ops and seal moved onto it; before,
  each read its key with the library's lenient decoder and checked the lenient
  equation, which accepts a key of small order (under the identity, `(R, S) = (B, 1)`
  verifies over every message), a second spelling of a key, and an `R` of small order.
  Claims, JWP groups and the key log, which held the rule in a copy of their own, call
  it too. It builds under the `ed25519` feature, the dependency alone.
- **One verifier in Python.** The claim, JWP, signer, HTTP and packaging verifiers share
  `claim-conformance/_ed25519.py`, the rule written from RFC 8032 with the standard
  library. `pack_verify.py` had checked signatures leniently.
- **Corpora.** Each receipt family's corpus carries the three signatures a verifier
  checking only the equation accepts — the identity as the key with `(R, S) = (B, 1)`,
  the identity spelled `y` = p + 1, and an `R` of small order under the base point —
  and every conforming verifier refuses them: new cases in jwp-conformance,
  signer-conformance and pack-conformance (`mutations/` and `receipts/`); claim- and
  http-conformance had them. No existing vector's expectation changed.

### The int_codec model and vectors: trained and coded on a generated signal

- The int_codec model and vectors are now trained and coded on a generated signal;
  earlier revisions used a third-party recording.
- `tools/wai_int_audio_export.py` reads no audio file. It writes its training signal
  and the clip it codes from a seed (16 kHz mono: tones, chirps, noise bursts and
  speech-like formant sweeps), in integer arithmetic, so a seed gives the same samples
  on every machine; `--signal-only` writes one with the Python standard library alone.
  Training is seeded from the same seed with deterministic algorithms and a fixed
  thread count: two runs gave byte-identical files. The seed and the command are in
  [`wai-rs/tests/vectors/int_audio/README.md`](wai-rs/tests/vectors/int_audio/README.md).
- Regenerated through their generators: `wai-rs/tests/vectors/int_audio/` (the model,
  both bitstream forms, `audio.json`, `sha256.json`); `int-payload-conformance`'s
  `vectors.json` and its thirteen `wia1_*` refusal fixtures; the staged-delivery case
  `audio_int_codec_1l` and `index.json`; the staged-measurement audio title,
  `goldens.json` and `sim/results.json`. The `WIA1` format, the capability and the
  model's shape are unchanged. `wai_int_payload_vectors` now reads the place of a
  `WIA1`'s first table from the payload's header for the two fixtures that change it, so
  they keep their meaning on tables of another length (on the earlier vector the
  offsets are the same).
- `tools/wai_audio_provenance_check.py`, run by CI: every committed audio vector is read
  to PCM and correlated, at every lag, with the reference recordings the CI host has;
  above 0.5 it fails, and a host with none skips it and says so.
- Every corpus states its content's provenance: a `**Provenance.**` line in each
  `*-conformance/README.md` (OER/2's settled corpus aside) and a row per directory in
  `wai-rs/tests/vectors/README.md`, required by `tests/corpus_provenance.rs`.
  CONTRIBUTING states the rule. Two vectors are flagged there as not generated and to
  be replaced: `int_video`'s `video_real` and `neural-conformance`'s `photo`.

### Staged tracks over MoQ: a track per layer, a receipts track per track, delivery timeouts per hop, chains that recover

- **Re-read against the drafts.** [moq-streaming-format](extensions/moq-streaming-format.md)
  names the revisions it was last read against, `draft-ietf-moq-transport-22`,
  `draft-ietf-moq-msf-01` and `draft-ietf-moq-cmsf-01`, and cites their sections. The
  re-read corrected: MSF's packaging values (`catalog` is a track name, not a value); what
  orders delivery (MOQT, subscriber priority first, then publisher priority, MOQT
  §5.1.2); how properties are carried (MOQT's track and object properties, of which MSF
  defines one); `SUBGROUP_DELIVERY_TIMEOUT`, which runs from the subgroup's FIN to its
  acknowledgement, not from an object's arrival; that MSF maps every object to a stream
  of its own, so a staged object never travels as a datagram; and that MOQT-22 replaced
  the Joining FETCH with fill fetch streams.
- **§4.1 Staged tracks.** A stage set is carried one track per layer: base tracks and a
  track for each layer index, each object one whole envelope. Each entry carries
  `wai_staged` (`set`, `layer`, `lead_ms`). A set's extent is its tracks, layers 0 to the
  highest with a track at each, so an MSF delta update that adds, removes or clones a
  track changes it alone; `lead_ms` is read only while the set is live. Tracks are named
  by namespace and name (a namespace inherited as MSF says), unique in a catalog, one
  namespace to a set; the check needs the catalog track's namespace, and a checker that
  does not know it refuses a set that mixes declared and inherited namespaces; `renderGroup` is an integer that fits 64 unsigned bits and
  `depends` an array of strings, each naming a track of the catalog in the entry's own
  namespace; a layer track depends on a track of every lower layer and on none of its
  own or above. A sink checks a set's entries in a fixed order and reports the first
  failure (`staged-malformed`, `staged-set-incomplete`, `staged-depends`,
  `staged-grouping`, `staged-lead`, `staged-receipts`), treating a failing set's layer
  tracks as absent. On a live track each group carries one object in a subgroup of its
  own. The base track's publisher priority is first among the set's (lower numbers first),
  though a subscriber's own priority comes before it. A layer object's
  `OBJECT_DELIVERY_TIMEOUT` is at most its slack to the last unit it covers, settable
  per object; a base's never shorter than the time to its last unit.
  Timeouts run per hop and never decide lateness; a sink's longer target latency cannot
  extend a publisher's timeout. A subscriber caps its stage, and its receipts, by
  subscription.
- **Receipts.** `wai_receipt_alias` (§4) is the `track_alias` a track's receipts sign:
  MOQT's Track Alias is chosen per session by each publisher on the path, so it is not
  that number past the first hop. Every alias in a catalog is an integer, compared as a
  number. `wai_receipts_track` names each media track's own **receipts track**
  (`wai-receipts`, at the media track's priority, no delivery timeout), whose group *g*
  carries the receipts of the media track's group *g*. No two tracks share an alias or a
  receipts track: a rendition of a layer, a clone among them, overrides both and has a
  chain of its own. Each receipt says which MOQT object it is about — `group_id` the
  MOQT Group ID, `object_id` the Object ID, the publisher's key and the track's alias —
  and a subscriber refuses one that does not (`moqt_mismatch`, a new jwp-receipts §3.3
  reason) before verifying it, holding nothing of it. A subscriber sets no delivery
  timeout on its subscription to a receipts track. A MOQT relay that carries receipts as
  opaque payload is not a jwp-receipts §4 relay; the subscriber verifies every receipt
  (jwp-receipts §4.1 amended).
- **Recovery.** A missing receipt is recovered by FETCH of its receipts-track group. A
  subscriber states its bound in counts: receipts held live for each chain (a time only
  beside it, never alone), held as history for each chain, and held across its chains.
  Past the live count the held receipts are refused (`parent_not_recovered`, a new
  jwp-receipts §3.3 reason) with their layers, the chain re-anchors at its next validly
  signed group, and the layers of the groups the anchor skips are refused. A subscriber
  that joins mid-stream declares a join and anchors at its first group. Each anchor
  leaves a gap, which a group linking to either end fills as history, in any order — the
  ascending order of a FETCH or a fill fetch stream included — and never as a fork;
  history waits under its own count. A group admitted before, offered again with the
  same bytes (MOQT can deliver one object twice, by a subscription and a fill), changes
  nothing (`accept again`), never `chain_break`, and its objects verify against it as
  against the head. jwp-receipts §3 step 3 defines holding,
  re-anchoring, joining and history, and says what an anchor does not prove; §6.1 says
  how anchors are reported and why the bounds hold across chains.
- **Other text.** staged-delivery §8, §9 and §13; jwp-receipts §4.4 and §5.9;
  http-delivery §10.2 and Appendix C. The registry lists `wai_staged`,
  `wai_receipt_alias`, `wai_receipts_track` and `wai-receipts` as wire identifiers.
- **Reference implementation.** `staged::moq` (feature `staged`): `StagedTrack`,
  `SetTrack::to_catalog_json`, `receipts_track_json`, `check_catalog` and
  `check_catalog_in`, `depends_for`, `live_object_rules_ok`, `priorities_ok`,
  `effective_timeout_ms`, `expired_at_hop`, `layer_timeout_ms`, `base_timeout_floor_ms`.
  `jwp_delivery::ReceiptChains` gains `with_hold`, `with_join`, `admit`, `expire`,
  `take_events` and `anchors`, and for MoQ `for_moq` (every count at most
  `MAX_HELD_GROUPS`) and `admit_on`; `HoldBound` states live, history and total counts
  (`from_json` refuses a statement without them); `ObjectReceipt::check_moqt`. Each
  group's signature is verified once, and a held group is offered again only when a
  group it links to is admitted (`verifications`, `reoffers`, `held_total` count them).
  Without recovery it behaves as before. `tests/staged_moq.rs` runs the corpus trace
  with signed receipts through `for_moq` under each subscriber's stated bound,
  per-track chains, dropped objects checked by inclusion and named by the relay's
  `delivery-gap` claims, and sealed `stage` claims.
- **Corpora.** `staged-conformance/moq/`: the set's MSF catalog; catalogs around §4.1's
  check, among them type cases, namespace inheritance under the catalog track's
  namespace, another and none known, MSF delta updates (a clone of a layer refused
  unless it overrides its alias and receipts track) and conversion to on demand; and a
  trace of the set through one relay to four subscribers, two layer objects dropped on
  timeout and one forwarded within its timeout and late, one subscriber's receipt lost,
  recovered by re-anchoring and the gap then fetched and filled as history, and one
  subscriber joining mid-stream. `rate_port.json`:
  the `rate` predictor beside the f64 interpolation of the layered-PCM ladder idea,
  equal wherever that form is exact, differing (pinned) where it rounds next to a half.
  `jwp-conformance/` gains recovery vectors — among them a gap fetched in ascending
  order after a re-anchor, a joiner's ascending fill under a live bound of 4, the
  history bound passed, the bound across chains, and a group delivered twice — appended
  after every existing one. `staged_verify.py` and `jwp_verify.py` check all of it independently;
  `chain_gap_differential.py` lets both chain verifiers recover.
- **joule-moq** (`joule-code/`): `staged` (the `wai_staged` member, the layer read from the
  envelope, the prototype `mrl_*` header compared and never obeyed), `ReceiptChains` per
  key and track, verifying each group's signature itself and knowing a group admitted
  before by its bytes, with the same recovery and bounds, and a `MockRelay` that forwards staged tracks on
  a caller-driven clock: per-hop timeouts, capping by subscription, one object to a live
  group, a receipts track per media track, `RelayStats` keyed by track, group and object,
  and send times past the clock refused. Its test `staged_tracks_drop_on_timeout_keep_receipts`
  runs the same trace.
- No existing envelope, receipt or claim byte changes. Two existing jwp-conformance
  vectors change their expected outcome: an exact replay of a group the chain admitted
  (`chain_head_predecessor_replayed`, and g1 in `chain_legacy_replay_loop`) is now
  `accept again`, not `chain_break`.

### The session claim: what a sink reconstructed

- **The `reconstruct-session` class** ([jwp-receipts](extensions/jwp-receipts.md)
  §2.4.12, the next number after `stage` and `delivery-gap`): a sink's statement of one
  session, about its entry object — the decoders it used, on which host and against
  which parameter sets; the path each unit took and why it switched; the presentations
  it made after their deadline; a commitment to every output it presented, by the root
  and leaf count of [staged-measurement](extensions/staged-measurement.md)'s output tree;
  the bytes it counted and received, and the objects it could not count; the classical
  rendition it did not need, by its `WAI2` table length; the `stage` claims it signed
  about the stage sets it observed; and, where metered, the decode figure of one
  completion with its acquisition class (`null` when unmetered, never 0). Its rules at
  §3.1 step 4, and step 6 against the presentations (the tree, first-use order of the
  decoder entries, paths, units, missed presentations, and re-decodes through decoders
  whose replicate class makes them byte-exact), the objects received (its entry object
  and every decoder entry's object among them), the envelopes received and the `stage`
  claims it names. Each rule
  and check has its reason in §3.1's table. Its capped lists are exact: the switches and
  the missed presentations are each listed in full up to the list's cap, and counted in
  full.
- **It restates no `stage` claim.** A `stage` claim states one set's layer objects, its
  units at each stage on time and its delivered stage; the session names it by link hash
  with the title's unit that is the set's unit 0, and commits to every output across the
  title, late ones included. A sink signs one for each epoch of a set — before it forgets
  the set's base, and at the end — and names them all. Step 6 binds the two both ways:
  each set's units were presented; the session presented, at each stage, no fewer units
  than a `stage` claim counts; no unit at a stage beyond the layers that the claims whose
  sets hold it — every epoch's, and every base's over the unit — list as `applied` or
  `partial` (a `late` or `refused` layer applied to no unit); and each claim's base and
  arrived objects are among the objects the session received. Nothing binds a `stage`
  claim to a session beyond its base and key: one an earlier session of the title signed
  can be named.
- **A decoder entry names the object it decoded**, and the envelope binds it. Its
  `object` is the content hash of the `WAI2` envelope whose rendition it decoded, or of
  the `WAI1` envelope; it is among the objects received (`session_source_not_received`).
  The envelopes a verifier holds — every `WAI2` envelope received, and any `WAI1` one —
  are among the objects received, read as plaintext envelopes
  (`session_envelope_unreadable`), and an entry whose object is one of them names what it
  gives: a rendition and that rendition's capability, or the `WAI1` manifest's capability
  or fallback (`session_source_mismatch`). A signer cannot relabel an entry whose envelope
  the verifier holds out of the re-decode check, by any route; which object an output came
  from remains its statement. An entry decodes one object, so a title's output-tree
  entries are (capability, pin, object) as well (staged-measurement Appendix C), and a
  session over a title still commits to the title's tree.
- **The classical rendition.** An envelope's classical rendition is the first selectable
  one whose capability the codec glossary links from an entry of era `classical` and
  whose media is the envelope's; WAI's own formats, neural codecs, records, parameter sets
  and text never qualify.
- **Authorities** (§7.5): a session needs `reconstruct-session`, not pinned-only. Its
  energy figure is a second item for confirmation, needing `reconstruct-session` and the
  energy authority of its class; a verifier reports the two side by side, so an origin
  can stand behind a session and not behind its figure. A session's subject's origin is
  not expected for it: a relying party names the origin it expects, if any.
- **Figure profiles** ([energy-measurement](extensions/energy-measurement.md) §6.1): the
  `moq-jwp:claim-link` row widens to a session whose `energy` is not `null`, so a
  `measurement` claim publishes the bracket behind a session's figure.
- **Delivery telemetry**'s `claim` member names a registered class now
  ([delivery-telemetry](extensions/delivery-telemetry.md) §2.1).
- **staged-measurement §8.6–§8.7**: a tool probes a power reader's rail at rest before it
  brackets a decode — absent, stale or live by energy-measurement §3.1's rule — and reads
  an energy counter as power by its successive differences across its wrap, ending the
  samples at an interval that cannot be a reading — one implying more power than a
  processor rail draws, one longer than ten sampling intervals, or a difference of more
  than half the counter's range (a counter reset); a tool that signs a session claim
  seals a figure only where its verdict is clean, beside the `measurement` claim on it,
  its burst run on a power of two of workers, which the session's host states as its
  `threads`, so the claim names no more of the machine than the session
  (jwp-receipts §6.7).
- **Reference implementation.** `claim_receipt::session` (the class),
  `session_recorder` (a sink's tally, each decoder named with the object it decodes; it
  verifies each `stage` claim as it records it — one for each epoch of a set — requires
  it signed by the key that finishes the session, and checks the sealed session against
  what it recorded and against those claims before it is returned),
  `key_log::Item::figure_of_claim`, `energy_report::rail_liveness`,
  `energy_burst::counter_power_samples` and `rapl_power_reader`,
  `energy_burst::{measure_on, power_of_two_workers}`, `stage_measure::run::completion`,
  `stage_measure::decode::{steps, prepare}` and `stage_measure::session`, and
  `wai_stage_measure bench`, which times each case of a title — one base's pinned decode
  or one layer's offer, its checks and hashing after the clock — over rotating rounds,
  compares every run's outputs with the evaluation's and aborts on a difference, records
  each run and the operator's statement of whether the machine was idle, meters the
  decode where a live rail allows, and with `--sign` seals the title's session and the
  report on its figure. A lib test replays every `staged-conformance/` case as a sink and
  requires the session it records, naming each epoch's `stage` claim, to seal. The
  output tree builds under `receipt_claims`; the switch reasons live in the ungated
  `delivery_terms`, which `telemetry` re-exports, so `receipt_claims` does not take
  `telemetry`.
- **Corpora.** `claim-conformance/` gains the session vectors (§5.11) — every reason the
  class adds, output trees of 1 to 33 leaves, every context, the classical rule across
  the registry's kinds of capability, the caps at their edges, a base forgotten and
  decoded again (`staged-conformance/`'s `record_budget`, its envelopes and each epoch's
  `stage` claim) and every route by which an entry could be relabelled — and
  `signer-conformance/` the session's authority and its figure's; `claim_verify.py`,
  `signer_verify.py` and `differential.py` read the class. `claim_verify.py` reads
  envelopes with `staged-conformance`'s strict readers, and `differential.py` compares the
  two implementations across both stage bounds and many epochs, the caps, malformed and
  encrypted `WAI1` and `WAI2` envelopes, entries naming their objects or not, and every
  registered capability as a classical candidate. Every vector already in both corpora is
  unchanged, and the claim corpus pins the vectors before the class by a digest computed
  before it was added. `staged-measure-conformance/goldens.json` restates each title's
  decoder entries with the envelope each decoded, and so its output root.
- **Wire.** No existing receipt, claim, pin or corpus vector changes bytes; a verifier of
  an earlier revision refuses the new class as `unknown_class`. A title's output root
  under staged-measurement Appendix C changes, its entries now naming their envelopes.

### Stage receipts: chains per key and track, objects verified by inclusion, `stage` and `delivery-gap` claims

- **One receipt chain per publisher key and track** ([jwp-receipts](extensions/jwp-receipts.md)
  §2.4.1, §3 step 3, §6.1). `group_id` and `parent_receipt_hash` run per
  (`signer_pubkey`, `track_alias`). The text disagreed with itself: the signing
  payload's comment and §6.1's last bullet said per track, §2.4.1 and §6.1's first
  bullets per key; all now say per key and track. Receipts issued with `track_alias`
  `0`, or chained per track, verify as before. Receipts a publisher chained per key
  across non-zero tracks, as the earlier §2.4.1 allowed, now fail at §3 step 3: each
  group whose parent is on another track names nothing its own chain holds
  (`parent_not_held`). A stage set carries each layer on a
  track of its own, and a subscriber verifies the tracks it subscribes to without the
  receipts of the ones it does not. A group that passes §3 steps 1–3 can head its chain
  whether or not its objects arrived, so a dropped object does not break its track's
  next link. §6.1's checks are now real in the reference verifier
  (`jwp_delivery::ReceiptChains`): the monotonic `group_id` guard per key and track
  (`group_id_not_increasing`) and the freshness bound on `group_end`
  (`group_not_fresh`), off by default for labelled groups.
- **A legacy link authenticates nothing** (§3 step 3, §6.1). The legacy payload signs
  neither `parent_receipt_hash` nor, through the link hash, anything that covers it, so
  a relay that rewrote one link — the first group's, to name the head — made the
  chain's whole history link again as new. The legacy payload does sign `group_id`, so a
  verifier now refuses a legacy group whose `group_id` does not exceed its chain head's,
  whatever its opt-in setting. A labelled replay fails its signature.
- **Forks are told from gaps** (§3 step 3). With a chain's head held, a `null` parent or a
  parent that is another group the verifier holds is `chain_break`; a parent naming
  nothing it holds is `parent_not_held`. A chain's live edge is the highest `group_id`
  of a group that verified and arrived inside the freshness window, linked or not; a
  group whose signed `group_id` is below it is history, exempt from the bound, so a
  group a relay delayed past the window no longer wedges the chain behind it, and a gap
  of any length fills in any order. Both payloads sign `group_id`, so no rewritten link
  earns the exemption. A previous group of another key or track
  is not held: a `null` parent beside one starts the chain, any other parent is
  `parent_not_held`. `group_chain_break` in `jwp-conformance/` now holds a second chain
  head; its earlier input, a parent that is not the previous group held, is
  `group_parent_not_the_previous` and is `parent_not_held`.
- **`inclusion-verified`** (§3, *Verified by inclusion*; §4.1, §4.4): an object receipt
  checked without the object's bytes — steps 1, 2, 3 and 5 — establishes that the
  publisher signed an object with that content hash at that position, and is never
  acceptance or delivery. A relay forwards the receipts of every object it does not
  forward, whatever the reason, and **MAY** say why in a `delivery-gap` claim (§4.2).
  `ObjectReceipt::verify_against(group, envelope: Option<&[u8]>)` returns
  `Evidence::{DeliveredVerified, InclusionVerified}`, which converts into staged
  delivery's `ReceiptEvidence`.
- **The `stage` class** (§2.4.10): what a party observed of a stage set, about its base —
  the stage requested, committed, delivered and presented, as points of staged
  delivery's vocabulary `urn:wai:staged-delivery:1` (registered in
  [energy-binding](extensions/energy-binding.md) §7.2; §2.4.9's vocabulary row names
  it); units by stage; the tier and stage hash of what was presented; every layer
  object with its units, bytes, status, evidence and staged-delivery §4.2 reason;
  conflicts, deadline misses and envelope bytes. Six rules at §3.1 step 4, in order:
  the points; the stage hash; the units; the objects (in order and within the set,
  except an object refused before its span was checked; evidence `delivered-verified`
  for an applied layer wherever a delivered stage is stated; bytes for every object
  that arrived and none for a dropped one, fewer in all than the set's; no layer applied
  twice to a unit; a partial object over at least two units); the units against the
  objects, per unit (an applied object of layer *s* lies within the spans of layer
  *s* − 1, a partial one meets them, and the units at stage *s* or above are no more
  than layer *s* can have reached, a partial object counting at most its span less
  one); and the delivered stage, which must be the
  one the objects give. Delivered and presented are separate measures: a set whose
  layers were all refused meets its commitment with stage 0 presented. A delivered stage
  may be stated with no commitment beside it, unlike §2.4.9's rule 3: the observer
  states what arrived. Whether a commitment was met is derived, never stated.
- **The `delivery-gap` class** (§2.4.11): a relay's statement that it did not forward
  the object at (`track_alias`, `group_id`, `object_id`) — `delivery-timeout`,
  `discretion` or `not-received` — checked at step 6 against the object receipt at
  those coordinates (`gap_subject_mismatch`).
- **Authorities** (§7.5): a `stage` claim needs `stage`, a `delivery-gap` claim
  `delivery-gap`; neither is pinned-only, and `delivery` grants neither.
- **staged-delivery §10** is split: §10.1 attestation of layers, which now names
  `delivered-verified` and `inclusion-verified` and the per-track chains, and §10.2
  claims. `staged::claim::StageObserver` states a sink's `stage` claim from what it
  observed: the layers that arrived, with the layer and span their own `refines`
  states, those known only by their receipts, and every report the sink gave, the
  latest deciding. Every event carries the sink's epoch for the base
  (`StageSink::base_epoch`, which counts the times the sink has forgotten it): a later
  epoch starts the observation over, keeping only the envelopes the sink still holds for
  the base, and a claim never spans two epochs. Every case of `staged-conformance/` is replayed through a
  sink and its observers, and each claim seals and states what the sink did. [http-delivery](extensions/http-delivery.md) §10.5 says what an
  abandoned refinement is: `inclusion-verified`, and `dropped` or `late` in a `stage`
  claim; [staged-measurement](extensions/staged-measurement.md) §10.2 says the same of
  an object its simulator drops.
- **Corpora.** `jwp-conformance/` gains the `chain` procedure — vector
  `v6_per_track_chain` (§5.9) and its neighbours: interleaved tracks, a track joined
  mid-chain, a link across tracks, forks, the guard and the freshness bound — and the
  `inclusion` procedure. `claim-conformance/` gains both classes (§5.10), every reason
  they add, and a context for a gap's object receipt; `signer-conformance/` gains the
  authorities they need. `jwp_verify.py`, `claim_verify.py` and `signer_verify.py`
  report every new outcome; `differential.py` generates `stage` bodies around every
  rule and gaps against receipts at and away from their coordinates, and
  `chain_differential.py` signs chains around §3 step 3 and §6.1 for both verifiers.
  Every vector already in the claim and key-discovery corpora is unchanged, and the
  claim corpus pins the ones before the new classes by a digest computed before they
  were added. In `jwp-conformance/`, `group_chain_break` alone changes, as the item on
  forks and gaps says.
- No receipt or claim byte of existing material changes; one existing vector,
  `group_chain_break`, is given a genuine chain break, as above.

### HTTP delivery (Draft)

- **The extension** [http-delivery](extensions/http-delivery.md) binds envelopes to
  HTTP, with no envelope byte and no manifest member changed:
  - the proposed media type `application/wai`, not registered;
  - `codecs` elements `wai1.`/`wai2.` that list the exhaustive union of the
    capabilities a variant may dispatch to or apply, with the floor every object of it
    offers carried separately (`X-WAI-FLOOR`, a DASH `SupplementalProperty`);
  - objects named by their receipt content hash (`<cid>.wai`), immutable, with no
    content coding and no `Vary`, and exactly one envelope each;
  - the `WHP1` head proof: the BLAKE3 chaining values that prove an envelope's head
    against its content hash without the rest of the object;
  - verified range fetch of one `WAI2` component that picks as whole-envelope
    selection does, and falls back to the whole object when a pick has no digest, a
    digest fails, or the plan needs most of the object;
  - receipt, key, grant, parameter-set and playlist sidecars; cache rules; priorities,
    deadlines and the carriage of enhancement companions, refinement components and
    layer objects ([staged-delivery](extensions/staged-delivery.md));
  - HLS and DASH adjacency for on-demand, live and low-latency live; a capability
    probe that never asks a user agent about a WAI element.
- **jwp-receipts**: §3 step 4 accepts the content hash a verified head proof
  establishes for an object fetched in part, and checks a receipt about a component's
  exploded envelope over the envelope the verifier builds. New §4.4 (objects not
  delivered) and §4.5 (an HTTP cache is not a relay). §1 and §7.9 point to the HTTP
  locations. moq-streaming-format §6 names the HTTP binding.
- **SPEC**: §2 and §7.1 name the HTTP object and the `WAI2` head; the
  `wai.video.manifest` row admits `render` and `extras` on a playlist receipt.
- **The registry's wire identifiers.** `codecs::registry::WIRE_IDENTIFIERS` lists the
  names the container, the two delivery bindings and key discovery put on the wire
  outside a manifest, each with its kind, defining section and status; `wai registry
  export` writes them to `registry/capabilities.json` (`wire_identifiers`) and the table
  in `REGISTRY.md`, and a library test checks each against the section it names.
- **The reference binding** (`wai-rs`, feature `http`): `http_binding::{head, head_proof,
  plan, session, codecs, origin}` — the head and §5.1, the `WHP1` proof built and checked
  with BLAKE3's tree, the root checked before any covered byte is read as a head; plans by
  priority class; a sans-IO fetch session with §7.1's response checks, component digests,
  the whole-object fallback and refinement deadlines; the `codecs` grammar, union, floor
  and support; an origin's headers and Range answers. `container::select_entries` is SPEC
  §7.1's walk with the digest test passed in, which `WaiMulti::pick` and a client holding
  only the head both run, and `companion_indices` / `refinement_indices` bind by entry.
  The `wai` tool gains `wai http publish|codecs|origin|vectors|bench`.
- **The corpus** [`http-conformance/`](http-conformance/): objects with their head proofs
  and ETag sidecars (a pinned learned image beside AVIF, JPEG XL and PNG of the one
  picture; AV1 at two rates with HDR10+ bound by id; a missing and a wrong digest; an AVC
  base with an LCEVC enhancement; a segment sequence; Appendix D's envelope; a head of
  several chunks) and cases for §4–§7. `wai_http_vectors verify` and the independent
  `http_verify.py` (standard library, written from the text, checked first against
  published BLAKE3 vectors) agree on every case. Falsification patches and Python
  mutations each must be caught, and a differential checks wai-rs's proofs of random
  envelopes in Python (`.github/workflows/wai-http.yml`).
- **Cache conformance** [`cache-conformance/`](cache-conformance/): a driver for §9's
  rules, run against the reference origin (which must pass), against four
  negative-control proxies (which must each be caught) and against labelled caches where
  an environment provides them; reports name a cache by label, version and configuration
  digest only, and CI uploads them.
- **Playlists and live** (http-delivery §8.6, §10.4, §11, Appendices G and H):
  - `hls_render::render_hls_with` renders on-demand, live and low-latency WAI media
    playlists (media and discontinuity sequences, a fixed target duration, parts with
    their server control, discontinuities, program date-times, `#WAI-LAYER` lines with
    `N` only, the open segment and its preload hint) and, under the legacy parameters
    with no extras, gives `render_hls`'s bytes; the module needs no dependency, and
    `wai_video_manifest` re-exports it under its old paths;
  - the playlist receipt (`ManifestReceipt`) carries `render` and `extras`; its
    signature binds a render digest only when a rendering is present, so every receipt
    sealed before verifies unchanged (one sealed by the earlier tree is pinned), and a
    rendering has one form; `seal_with`, the receipt file's strict JSON form and
    `mr/<manifest_hash>.json` are new, and a live playlist's revisions chain;
  - `http_binding::hls` reads media playlists, holds a WAI playlist adjacent to its
    classical one (durations to the millisecond, rounded half up, read exactly), renders
    the WAI media and multivariant playlists (bandwidths computed from the objects),
    and keeps the classical master to its one session-data line (`classical_changed`);
    `http_binding::dash` writes `wai.mpd` on the classical timeline;
    `http_binding::playlist_receipt` seals and checks playlists and live chains
    (`playlist_receipt_malformed`, `_mismatch`, `_invalid`, `_unchained`;
    `playlist_malformed`);
  - `wai http hls|mpd|seal-playlist|check-playlist`; the origin serves playlists to be
    revalidated on every use;
  - the corpus gains a staged pair with a layer over its second segment (O11), a
    low-latency set of two segments of twelve parts each, every part its own object
    (O12), a player fixture's WAI segments (O13) beside a playable classical
    presentation, the WAI playlists beside their classical ones, their receipts (the
    live revisions chained) and the published classical master. O10 was already the
    layer set, so these took the next numbers.
- **Adaptation** (http-delivery Appendix F): `http_binding::abr`'s four reference
  policies (throughput, buffer, utility over buffer, look-ahead) in integers, the budget
  rule that never uses an unlabelled figure, an entry order the session applies on both
  paths (the selector's order now sees the verified head), and layer depth; seeded
  vectors per policy, the integer log, entry orders and depths, which `http_verify.py`
  reproduces.
- **Playlists and adaptation, after review** (http-delivery §8.6, §11, Appendices F, H
  and I):
  - a live client walks back to the head it holds across revisions it skipped: every
    receipt is also served at `mr/<receipt_hash>.json`, the head fetched again is
    unchanged, and a walk is bounded (`playlist_receipt_parent_missing`,
    `playlist_receipt_walk_exceeded`); no revision lists another cid for an alias or a
    media sequence number an earlier one bound (`playlist_listing_rebound`);
  - Ed25519 is verified strictly (`wai::sig::verify_strict`: a canonical key of large
    order, `S` < L, `R` of large order compared byte for byte), by the playlist receipt
    and `http_verify.py`;
  - DASH adjacency is checked over an XML subset (Appendix I); `wai.mpd` takes the
    classical MPD's period, timescale, `t` and `d` and leaves a gap out; MPD values are
    escaped and checked as XML;
  - adaptation states have stated bounds (`abr_state_out_of_bounds`) and are computed
    exactly within them; under a budget only a labelled option within it is chosen;
  - a gap segment carries no parts or layer lines; the session-data line has one place
    in the classical master; a receipt file holds no member twice, no lone surrogate
    and no number past 2^53 − 1, a grant of no joule ceiling written `null`;
    `seal-playlist --grant`; the origin answers a far-future `_HLS_msn` with 400.
  - an MPD is held as written: timelines in runs, never expanding `@r`, compared run by
    run, at most 256 per document, and an element's attributes read in linear time
    (Appendix I);
  - a budget never raises the policy's choice: with nothing at or below it within the
    budget, option 0 and the budget `unmet` (replacing `floor`);
  - past the walk bound, a revision ahead of the head re-anchors the chain
    (`reanchored`), keeping its bindings, the revisions between unsigned; bindings key a
    line by its URL resolved against the playlist's (RFC 3986), a gap never withdraws a
    listed segment, and a binding is kept while one of the last 256 revisions lists it.
- **What review found, fixed** (http-delivery §3.1, §4, §7, §7.1, §10, §13, Appendix D):
  - a layer object (a `WAI1` whose capability is refinement-only) is never dispatched:
    it is fetched whole at refinement urgency, verified against its cid and handed to
    staged delivery, and joins a union after `+` through `refines.base`
    (`not_a_layer`, `codecs_layer_unbound`, `layer_tag_malformed`). A `WAI1` whose
    capability is a payload is an ordinary object even when it carries `refines`: it is
    dispatched, described by its capability, and `not_a_layer` when listed as a layer.
    A layer is declared only in the `alt` of its own base's capability;
  - a refinement never moves its base: only base-class bytes count toward the 90 %
    rule, a refinement is always requested on its own, a refinement with no digest or a
    failed one is dropped alone on the range path and on the whole path alike, and a
    refused refinement response abandons only it;
  - a client presents a base as soon as it verifies and applies its refinements as they
    arrive; a refinement's deadline runs from that moment on the caller's clock
    (`FetchSession::base_ready`, `tick`), and `wai http bench` reports the time to the
    verified base apart from the time to the complete fetch;
  - after a component mismatch every byte past `covered_len` is discarded before the
    rest is fetched again; a base request never answered is `base_unanswered`;
  - `Content-Range` numbers are ASCII digits only; an element beginning `wai` in any
    case must be a well-formed WAI element; an empty element or a capability named
    twice is malformed; a `WAI1` naming a record beside a payload fallback is described
    by its fallback; a `codecs` value with non-ASCII text, and a head proof whose lengths
    reach the end of 64 bits, are refused rather than panicking;
  - the reference origin caps what it reads of a request head (16 KiB); gives the whole
    head a deadline from the connection's accept (default 10 s, `--read-timeout-ms`),
    answered by 408 however slowly its bytes trickle in; bounds each connection's whole
    life from its accept, writing included (default 60 s, `--max-connection-ms`), so a
    peer reading a large response slowly is cut off then; serves at most 64 connections
    at once (`--max-connections`), answering 503 to one more, with a cap per peer
    address only when `--max-per-peer` is given (it listens on 127.0.0.1, where every
    client has one address, and a default cap would let a few reconnecting clients shut
    the rest out); decodes `/__hits` as bytes; and refuses `\`, `:` and NUL in a path.
    The wire-identifier test reads each name in its section's body.
  - a verified base is decided: a component that verified is never read again, and of a
    refinement's response only the range requested is kept, so a refinement answered
    `200` with the whole object, a byte changed, never reaches the base. Every response
    carries the caller's clock (`FetchSession::on_response(…, now_ms)`), the base is
    stamped with the time of the response that verified it, and the deadline's
    arithmetic saturates. The corpus's `fetch_timed` cases run the session on a clock;
    `http_verify.py` agrees on them.
  The corpus gains a case for each, and `http_verify.py` agrees with wai-rs on them and
  on a seeded differential outside the corpus, in which a refusal must match by its
  code and each family reports how many cases it generated and how many were refused.
- **Receipts, keys, grants and parameter sets** (http-delivery §8.1–§8.5, §10.5):
  - a receipt file `rcpt/<h>.json` holds one group receipt (in the labelled payload) or
    claim in JCS, at most 1 MiB, read as jwp-receipts reads one and refused unless its
    link hash is its name (`receipt_name_mismatch`, `receipt_not_served`,
    `receipt_too_large`); a legacy group, whose name does not cover its parent, is
    refused (`receipt_legacy_group`);
  - an index, at most 4 MiB with at most 256 delivery items and 256 claims
    (`receipt_index_too_large`), is read item by item (`receipt_index_malformed`), each
    receipt file fetched once per index, at most 1024 in all
    (`receipt_fetch_budget_exhausted`) and 64 MiB read in all, each file read no further
    than what remains of that (`receipt_fetch_bytes_exhausted`), each cap with its reason
    stated: each delivery item's group goes through
    jwp-receipts §3 against the client's chains, a missing parent fetched by its link up
    to 64 of them, never more, and never skipped (`parent_not_held`, with where the walk
    stopped), a joining client anchoring instead; its object receipt through steps 4
    and 5 against what the client holds — the hash recomputed over the envelope, the cid
    a head proof established (§8.3), or nothing, which is `inclusion-verified` and gives
    no property reading (§10.5); each claim by its subject against the items that
    passed, its parent fetched by its link, then §3.1 step 6 against what those items
    deliver and the receipts it names (`figure_not_held`, `context_not_held` where the
    client does not hold them; an energy claim naming an OER receipt, which this revision
    does not carry over HTTP, is `context_not_held`), a declaration it names read as
    §8.1 reads every file fetched by name before step 6 compares it, a recovery record
    checked with the client's matchers whether or not it holds the object, and a soft
    binding of another object attesting the ancestor only through §3.2's recovery,
    under the held publisher key that signed it;
  - a key log anchors only from the origin's own location over HTTPS, answered `200`
    with no redirect (`key_log_not_from_origin`), from a retrieval the verifier made
    itself, asking with `Cache-Control: no-cache`; a key log is read to 1 MiB and a longer
    one refused, anchored or cached (`key_log_too_large`); a log through any cache adds revisions
    and never moves the anchor, an older copy reported (`key_log_rolled_back`); a grants
    file is a locator (`grants_file_malformed`); a key status is served for a quarter of
    the head's `status_max_age_s` at most; a module grant is §7.9's form, and a JCP grant
    is carried only by reference in the receipt that names it;
  - a parameter set served by its digest is held only by its pin and §7.9
    (`pin::verify_bytes_with`, `int_payload::resolve`), never by its name.
  - `wai-rs`: `http_binding::receipts` and `http_binding::keys`;
    `jwp_hash::EstablishedHash`, built only from bytes or by a fetch session whose head
    proof verified (`FetchSession::established`); `ObjectReceipt::accept_hash` and
    `verify_against_hash`, `WaiReceiptProperty::check_against_object_hash` and
    `ReceiptChains::verify_admitted_object`, which the existing checks call;
    `energy_figure::HeldObject`, an object receipt's figure for step 6; `wai http
    receipts`, which fetches under a deadline and the caps above, follows no redirect,
    and with `--key-origin` retrieves the key log from the origin itself over TLS
    (the `cli_tls` feature: a TLS client validating against the operating system's
    trust store; SNI and the certificate check name the origin's host, port 443 by
    default, each address the host resolves to tried in turn), every step within one
    deadline however slowly a peer answers; a build
    without it refuses to anchor rather than fetch the log over plaintext, and the
    test-only `--test-plaintext-connect` anchors with its own path, `test-plaintext`,
    wherever the anchor or a confirmation is reported, as anchoring nothing a verifier
    may rely on. A group or
    object receipt whose `kind` is not a string is `wrong_kind`
    (jwp-receipts §3 step 1), as the independent verifier already read it.
  - The corpus serves receipts, indexes, a key log, a status, a grants file and a set;
    `http_receipts.py` checks them for `http_verify.py` through the jwp-receipts, claim
    and key-discovery verifiers, and the differential adds receipt files, indexes,
    grants files and key logs.
- No envelope byte or receipt byte changes for existing material.

### Packaging: objects, the stage index, the floor, catalogs and the package receipt

- **Objects** ([packaging](extensions/packaging.md) §5): each one envelope, named by its
  receipt content hash, with no manifest member naming a position, so a repeated piece
  is one object. The packer's base encode calls `choose_q` per piece.
- **The stage index** `WSIX` (§6, `wai-rs` `stage_index`, feature `video_package`): per
  component, every object of the package by content hash, its lengths, role, layer,
  the tier and decoded-output digest of the prefix it completes, its SSE, and the class
  plan; its digest and a Merkle stage root; a named reason for every rule a verifier
  applies to an index and to the objects it lists.
- **The floor** (§7): a supplied AV1 floor, checked against the keyframe plan
  (`wai-pack plan --out`) and the listing the policy pins, written as floor objects
  that carry `derived_from` and as CMAF files (`init.mp4` with `av01`/`av1C`, one
  fragment per piece).
- **SPEC §7.1, *Derived renditions***: `derived_from` on a manifest or a rendition
  entry states that a payload presents another object and was made from it; an
  OPTIONAL member under the MINOR rule that never breaks a parse (§8).
- **Catalogs** (§8): on http-delivery's Mode N, a classical master and floor playlist
  that name nothing of WAI; per class, its master (with the one SESSION-DATA line
  http-delivery allows), its WAI playlist with the codecs union and floor, and a WAI
  media playlist adjacent to the floor's, with `#EXT-X-GAP` for a piece the class
  takes as floor only and a `#WAI-LAYER` line for each refinement layer it lists (no
  package of this revision carries one: the packer refuses refinement rungs, §11.4);
  the classical and per-class DASH MPDs; the MoQ catalog; and a stage view per class.
- **The version-2 package receipt** (§9, `wai-rs` `package`): binds the policy digest
  and each component's index by digest and stage root, with entry, object and byte
  counts every verifier recomputes; RFC 8785 JSON both ways; a measurement claim beside
  a metered figure. The version-1 receipt gains the same JSON form and keeps its
  bytes. Both share the figure profile `wai:video-package-receipt-id`.
- **`wai-pack pack`, `verify` and `listing`**, and the corpus's golden package
  `video-base-01`, its index, object and receipt mutations with their reasons, and
  version-1 receipts as JSON. The registry lists packaging's wire identifiers.
- No envelope byte or receipt byte changes for existing material.

### Packaging: a declared policy, its digest, and pieces cut at scene changes

- **The extension** [packaging](extensions/packaging.md), first sections: a package is
  made under one TOML policy, `wai.pack.policy/1` (§3), read in a fixed order with
  every refusal named: value kinds a canonical JSON cannot carry exactly (floats,
  dates, negatives, integers past `2^53 − 1`), then the schema, every default written
  in, then the constraints across members (ladders and their refinement chains,
  classes, rules, deadlines against layer lags). The effective policy's digest is
  BLAKE3 over its RFC 8785 form, so two spellings of one policy have one digest.
- **Pieces** (§4): `wai.pack.scenecut/1` cuts a video component at scene changes by
  integer histogram distance, with a bounded lookahead whose return and revisit
  clauses keep a short flash from cutting; a decision reads no frame past the
  lookahead, so live and on-demand packing cut alike. Pieces are bounded by `min` and
  `max`, aligned across components to the lead, and classified static, synthetic,
  high-motion or natural. Y4M and WAV sources convert by stated integer rules.
- **`wai-pack`**, the reference packager, as far as a plan (`wai-pack policy`,
  `wai-pack plan`), with float arithmetic denied for the whole crate, and the corpus
  [`pack-conformance/`](pack-conformance/) with a second implementation in Python.
- **What a packager may emit** (§3.6): only capabilities SPEC §5 registers as payloads
  with a pinned format, in a role each can play, and under `royalty_free_only` only
  royalty-free ones; every refusal names the capability (`capability-unknown`,
  `capability-not-payload`, `capability-candidate`, `capability-role`,
  `license-unclassified`, `license-refused`).
- **One integer video encoder** (§5.5): `int_video_enc` codes RGB8 frames as the
  `WIV1` keyframe-0 payload SPEC §5 pins to `wai.video.int_motion`, with integer block
  motion search (whole- or half-pel) over the decoder's own prediction, residuals
  quantized in a closed loop, and one residual table for the clip; it states the SSE
  of its reconstruction. `int_payload::encode_wiv1_zero_motion` is its zero-motion
  case and writes the bytes it wrote before. The corpus holds its payloads for every
  piece of every fixture, decoded independently to the same frames.
- **`wai::jcs`**: the canonical JSON form moves out of `claim_receipt` into a module
  of its own, compiled under every feature set, unchanged in behaviour.
- **Choosing the step** (§5.6), specified and built ahead of the stage that writes
  objects: the largest of `q_candidates` (read as a set) whose SSE per sample, in
  thousandths, is at most `max_sse_px_milli`, else the smallest
  (`wai_pack::encode::choose_q`), with goldens recomputed independently.
- **Reading what a packager is given**: a policy is read from its bytes (UTF-8, no
  byte-order mark, nesting at most 32 deep, `policy-toml` otherwise), ids are unique
  across every kind, a cut window holds at least 3 scores, a WAV block alignment is
  compared without wrapping, and no declared size sizes a buffer; the segmenter's state
  is bounded by its lookahead and window. The encoder codes in two passes and states
  what it holds (`int_video_enc::working_bytes`), and `rans::RansReverseEncoder` codes
  values last-first into the same bytes.

### Staged delivery in the browser, a base writer, a trace and a bench

- **A base is bounded before it is decoded, and its chain from the start**
  ([staged-delivery](extensions/staged-delivery.md) §4.5, §4.2, §14). A base whose units'
  output, with 32 bytes for each stage its kept record holds, passes the chain budget is
  not added: its envelope is reported `chain-over-limit`, no chain is kept, its kept
  record is untouched, and its layers stay held. Where a base's capability declares a
  cost before decoding (the integer payloads SPEC §5 pins), a sink applies its cost
  limit to that declared cost first, and SPEC §4 treats a base past it as one it does
  not decode, moving on to the fallback; a capability that declares no cost
  (`wai.neural.int_synth`, `wai.neural.int_mlicpp`) decodes under its format's caps. The corpus gains `base_over_budget`;
  `staged_verify.py` implements the rule independently.
- **The sink in the browser.** `wai-web` binds `wai::staged::StageSink` for JS
  (`src/stage.rs`). `StageBase.decode` decodes a base envelope by SPEC §4 through the
  integer path, with its pin verified against the parameter set the page holds in
  memory and under the page's cost limit, which is a required argument (`null` for the
  format's caps only; it bounds a base whose capability declares a cost);
  `StageBase.fromStage0` takes a base the page decoded elsewhere, which is **claimed**:
  its capability must be the manifest's or its fallback and one SPEC §4 would take (an
  inert envelope is refused), its tier is never above what the capability gives an
  unpinned base, and every report, unit and stage hash of its chain says `claimed`. A base's facts cannot
  be raised from JS (no `viaFallback` setter; `attested` only rises), and `addBase`
  takes the base. `StageSink` adds bases, offers layers, expires holds, removes, attests
  and forgets bases, and reads back every unit's stage, producer, frozen flag, tier,
  shape, output, stage hash (with its `claimed` mark) and squared error against a
  reference. `stageLayerInfo` reads a layer's `refines`, `WIR1` header, declared cost and
  byte split (framing, manifest, payload), and a base's declared decode cost, before
  either is offered. Reports and units come back in the corpus's own form. Media reach
  the page as `WAI1` bytes only.
- **The browser equals the native sink.** `wai-web/demo/verify_staged_wasm.mjs`
  replays every case of `staged-conformance/` through the wasm32 build: each integer
  base is decoded in wasm32 and must give its `stage0.bin`, its tier and its fallback
  flag, and every report, final unit, stage hash and clip hash must equal what the
  native sink wrote. `wai-web-wasm.yml` runs it and watches the corpus; the crate's
  native tests replay the corpus through the binding's pure-Rust half as well.
- **One base path.** `staged::decode_base` runs SPEC §4 for a base through
  `Wai::dispatch_verified`: its capability, then its fallback, each only where it is
  dispatchable, so a refinement-only or companion-only capability makes the envelope
  inert (`missing_capability`) and never reaches its fallback. It takes the sink's cost
  limit (checked before any symbol is decoded or output allocated, `cost_over_limit`)
  and its key log (SPEC §3.1 check 5 over a pin and over an unpinned set; none skips
  it). `decode_base_via` asks for one of the two capabilities and refuses any other
  (`not-offered`). The refusals come in a stated order: `not-wai1`, then `encrypted`
  (SPEC §6.4 before §4), then `not-offered`, then §4's; a parameter set is read only
  for a capability that reads one. Its own tokens are staged-delivery's kebab-case sink
  vocabulary; SPEC and payload codes pass through in snake_case. A pin counts toward
  the base's tier only where it resolved. A
  one-file set pinned without a form offers every held file to the pin, the member the
  capability reads first. The CLI and the browser share this path.
- **A keyframe-0 base writer.** `int_payload::encode_wiv1_zero_motion` codes RGB8
  frames as a `WIV1` keyframe-0 payload (zero motion, one block per frame, a closed
  group of pictures): the reference encoder the corpus generator used, now in the
  library. The corpus regenerates byte for byte through it. It refuses a clip past the
  format's caps before allocating (`too_large`). `wai stage base` writes a
  `wai.video.int_motion` base with it from raw frames, with its dimensions multiplied
  checked, and decodes it back before writing.
- **A trace to check against.** `wai stage apply` is now a sink replay: envelopes are
  taken in order, routed by content (one with `refines`, or a refinement-only
  capability, is a layer; any other is a base, decoded by SPEC §4), and a base given
  after its layers releases the layers held for it. `--json` prints each step's reports
  in the corpus's form (bytes that are not a `WAI1` this sink reads, and a base it
  cannot decode, are the step's error by their code, and the replay goes on; a base
  read with `--stage0` is claimed, and its step, units, chain's reports and final entry
  say so), and every stage a unit reached in the step, in the
  order the sink applied the layers (`StageSink::set_trace`), so a batch that takes a
  unit up several stages lists each; every stage hash a delivery in that order reaches
  is in it. A test replays the corpus through it and requires the cases' reports and
  final units.
- **`wai_stage_bench`.** Measures one staged set natively: per stage, the envelopes'
  bytes (framing, manifest and payload, and cumulative), PSNR or SNR per unit and over
  the clip from the exact integer squared error against the source at the units' final
  shape (a unit of another shape has no figure at that stage), the time each stage took
  (median and quartiles over at least 20 runs), the direct comparator (one layer per
  base at the stage's step, straight from stage 0, or why it does not apply) and every
  unit's stage hash. Energy is reported as unmetered.

### Staged delivery: refinement layers, base admission, and WAI2 refinement components

- **The extension** [staged-delivery](extensions/staged-delivery.md): a stage set is a
  base and layers, each layer its own `WAI1` envelope naming, by content hash, the
  base and the producer each unit it covers is expected to have (`refines`: `base`,
  `prev` as one producer or runs, `layer`, `span`). A sink holds per unit a stage, a
  producer and a tier, and applies a layer per unit by one algorithm with a closed
  list of reasons (§4.1, §4.2): every envelope-scoped step (well-formed `refines`, a
  refinement capability with no fallback, the base present, not through a fallback,
  admitted, attested, a well-formed payload within the sink's cost limit and its chain
  budget) before the unit steps (conflict, readiness, geometry, deadline). Layers a unit is not ready for are held
  to a declared bound and released when its stage advances; layers released together
  form a batch, so two conflicting layers released together are both refused and the
  unit frozen, whatever their order. A layer that differs from one that already passed
  for the same unit freezes it at its stage. A unit's tier is the weakest of its base
  and layers, `entropy-consistency` does not pass through a layer, and a stage hash is
  stated only at `decode-equivalence`. Deferral, live deadlines (a sink never delays a
  unit for a layer, and counts the units presented below the stage their layers would
  have reached on time), attestation (a receipt verified without the layer's bytes
  never authorizes an application) and byte counting (envelope bytes; receipts,
  framing and parameter sets separately; never JSON or base64) are stated.
- **Resource bounds** (§4.5, §14, A.5): a sink declares a chain budget per base, over
  its units' output, 32 bytes per stage, and the payload and declared working bytes of
  every layer it retains with one unit's output and 32 bytes for each unit it covers
  and has not been applied to; a layer that would pass it is refused
  (`chain-over-limit`) before any stream is decoded, and a layer no unit is held for is
  dropped. The declared working bytes count the stream index. A sink's work per layer
  is linear in its payload and declared symbols up to a logarithmic factor in its
  units, and §14 states how. A sink that drops a base's chain discards the layers held
  for it (`base-absent`) and keeps each unit's record that is not empty (its
  producers, a pending producer, its frozen flag), within a declared record budget
  whose oldest bases are forgotten first; the base decoded again advances only along
  that record, and a frozen unit replays only its own record, in order.
  Decoding a base again keeps its chain, and a receipt verified for the base, before or
  after it was decoded, makes it attested from then on, across dropped chains; a layer held for its base and offered
  again keeps the strongest receipt evidence, and over an attested base counts as
  arriving when its receipt was verified; a set split across bases shares one live
  clock (`first_unit`).
- **The residual payload `WIR1`** (Appendix A): one rANS stream per unit;
  `clamp(pred + r·q)` with the identity, up2 (the half-pel arithmetic, odd sizes
  included) and rate (linear interpolation, `rdiv`'s halves away from zero, channels
  interleaved) predictors. It follows the integer payloads' rules: a parse in a
  stated order with the closed list of their codes, caps, a declared cost a sink reads
  before decoding, refusal at the first symbol a stream runs out on, canonical stream
  ends; and it reads no parameter set.
- **Refinement capabilities** (SPEC §5, *Refinement capabilities*):
  `wai.image.int_refine`, `wai.video.int_refine`, `wai.audio.int_refine` and
  `wai.splat.int_refine`, each refinement-only, royalty-free, `StandardDefined` and
  exact given the base it refines. SPEC §8 adds the refinement-only role; SPEC §4 never
  dispatches one, and a sink that does not implement the extension finds it
  unsupported. The registry gains the `refinement` kind and role, and the export is
  regenerated. *Envelopes this makes non-conforming* (they still parse): a `WAI1`
  carrying a refinement capability without `refines`, with a fallback, or as a
  fallback; `refines` on any other capability; a refinement capability as a `WAI2`
  rendition or companion; a companion bound to one.
- **Base admission is one table** (SPEC §5, *Base admission*;
  `codecs::BASE_ADMISSION`): the bases of each refinement and of the LCEVC enhancement
  layer, whose bases are unchanged. `enhancement_bases`, `companion_applies_to` and
  `admits` read it.
- **WAI2 refinement components** (SPEC §7.1), on the component identity (`id`,
  `sha256`, `prior`; no other digest is defined): an entry with
  `role: "refinement"` carries `refines` naming entries by `id`, follows every
  selectable entry, is never selected, and is checked by `check_component_roles` in a
  stated order. *Exploding a component* writes its `WAI1` envelope, from
  `envelope_manifest` verbatim or from the `WAI2` manifest's members as written, with
  ids replaced by the content hashes of the named entries' exploded envelopes, so a
  component and the layer carried alone are the same object; `WaiMulti::pack` packs
  `WAI1` envelopes with their manifests verbatim, and they explode back to the same
  bytes; it refuses an envelope it could not give back (bytes around a manifest's JSON
  value or after a payload, or a manifest that inside a `WAI2` passes the depth limit)
  and checks its output before returning it. Each entry explodes on its own, so one
  that cannot leaves the others intact. A refinement's `envelope_manifest` names no
  fallback. A malformed `refines` or a mismatched `envelope_manifest` does not make a
  manifest unreadable.
- **SPEC §2**: member names are unique within each object of a manifest, at any depth,
  and an integer member is written as an integer (`-0`, `1.0` and `1e0` are not); both
  readers refuse a manifest that breaks either. `span: null` is `span` absent.
- **Reference implementation**: `staged::{StageChain, StageSink, wir1, predict,
  wir1_arith, tier, schedule}` (feature `staged`); `container::{Refines, Prev}`;
  `WaiMulti::{explode, explode_each, explode_all, pack, check_envelope_manifests,
  refinements_of}`; `container::check_unique_members`; `wai stage
  refine|apply|inspect|pack|explode` (`refine --out-dims` gives a layer's shape);
  `StageSink::{remove_base, forget_base, mark_attested}`; `wai wrap` refuses a refinement
  capability, and `wai inspect` shows `refines`. `RansEncoder` writes the bypass
  escape, and `int_entropy::hist_to_cdf` builds a table from a histogram in integer
  arithmetic. In `ternary::rans_encode_trits`, the encode call sat inside
  `debug_assert!`, so a release build coded no symbol; it now runs in every build.
- **Conformance**: the corpus [staged-conformance](staged-conformance/) (image, video,
  audio and splat sets; cases for every reason of §4.2 and every `WIR1` refusal code,
  in the order the steps check them, and for §3's rules, holds and their bound,
  batches, frozen units, the chain budget, a base decoded again, receipt evidence and
  live deadlines, one clock across bases included; `WAI2` cases for each refinement
  rule of SPEC §7.1, for exploding entry by entry and for what a packer refuses; the
  live schedule), written and verified by `wai_staged_vectors` and verified
  independently by `staged_verify.py`, which shares no code with `wai-rs`; the pure-
  Python BLAKE3 moves to `claim-conformance/_blake3.py`, shared by both Python
  verifiers. Cross-architecture goldens pin every stage of an image, a video and an
  audio set; Kani proves, over bounded windows, the arithmetic step, the up2 bounds,
  `rdiv` and the rate indices. `rdiv` and the rate indices are computed in checked
  64-bit arithmetic.

### Staged-delivery measurement: counted bytes, integer quality, the output tree

- **The extension** [staged-measurement](extensions/staged-measurement.md): how a
  figure about staged delivery is measured and published. The counted bytes of an
  object are its envelope to the byte (`12 + man_len + P` for `WAI1`,
  `8 + man_len + 2 + 8n + Σ P` for `WAI2`), and an object is countable only where WAI
  checks the format of every payload it carries: the pinned integer formats and `WIR1`,
  by the parse a sink runs, for the capability and for a `WAI1` fallback with a pinned
  format. Any other capability, a classical codestream or an unregistered string among
  them, is not countable (`no-format-check`), and neither is a `WAI2` component whose
  `sha256` is not its payload's, nor an encrypted payload; no estimate stands in for a
  figure that is not countable. A figure carries the framing label `registered`. Quality is integers per
  presentation and per prefix, against the source and against the final prefix: SSE
  over RGB and over an integer BT.601 luma, a Q32 structural similarity over clipped
  8×8 windows at stride 4 with an edge-aligned last window on each axis, so every pixel
  is in a window (bounded in 128-bit integers), PCM SSE and reference energy,
  and tensor SSE and largest difference, each pooled over units; decibels are display
  only. First presentable, layering overhead as two rate–quality points (never a
  ratio), delivery simulation, decode time and energy (only a clean attribution verdict is a figure; a rail
  that is not live is unmetered), deduplication, dated records, what CI asserts, and a
  media-class matrix are stated. The output tree (Appendix C) commits to every
  presentation, one SHA-256 leaf per (unit, layers), under RFC 6962's tree hash, by its
  root and leaf count together, with audit paths checked as RFC 9162 checks them under
  the committed count (`output_tree::OutputCommitment`).
- **Measured titles** (Appendix D): a staged set over a staged-conformance base, read
  where it is, with only what the stage model does not carry beside it (unit duration,
  capture times, the source, single-layer references, the fixture tool's own
  reconstruction digests), and a content identity over the measured files.
- **Reference implementation.** `wai-rs` feature `stage_measure` (pure integer, no
  ONNX, wasm32-clean; CI builds it for wasm32): `measure_framing`, `output_tree`,
  `stage_measure::{title, decode, run}`, which decodes each base through the sink's own
  base decode, `staged::decode_base_pinned` (which also gives the pin it verified),
  under the cost limit, and applies the layers through `staged::StageSink`
  under the chain budget, handing each stage's output over as it is reached. The tools'
  stated limits are `staged::TOOL_DEFAULT_LIMIT` and `staged::TOOL_DEFAULT_CHAIN_BUDGET`
  (staged-measurement Appendix D states each figure and why); `wai stage` (`apply`,
  `refine`, `base`), `wai_stage_measure eval` and `wai_stage_bench` decode every base
  under them, hold every layer and chain to them, and take
  `--limit OUT,SYMBOLS,WORKING,MACS`. A title's `limits` lower a limit and never raise
  one: each figure is the lower of the title's and the tool's.
  The corpus's `refusals/` hold titles a tool must refuse, with the code: a base past its
  chain budget, the two declared-clip bound vectors (refused before anything is
  allocated for them; a counting-allocator test shows it), the 1 GiB one again under a
  title that states limits of 2^40, two inert envelopes, and a layer past its title's
  cost limit. `tests/vectors/staged_bounds/` gains a base whose up2 layer declares
  67,092,480 output bytes, which `wai stage apply` and `wai_stage_bench` refuse with no
  flag. Always on:
  `quality` and `content_dedup` (the package receipt's dedup figures now go through
  it; a map that gives one content hash two lengths no longer verifies); native only:
  `bench_env`, the machine record, which `wai_decode_bench` now prints from. The
  `wai_stage_measure` tool evaluates a title (`eval`), verifies the corpus (`verify`)
  and writes it (`gen-fixtures`, `generate-goldens`).
- **The corpus** [`staged-measure-conformance/`](staged-measure-conformance/): a
  picture, a clip, an audio clip and a splat attribute tensor, each with its figures in
  `goldens.json`; `wai_stage_measure verify` and an independent Python verifier,
  `measure_verify.py`, both reproduce every figure, and CI runs both.
- **The delivery simulator** (§10, Appendix B; `stage_measure::sim`,
  `wai_stage_measure sim`): a discrete-event model in integer microseconds of a sender
  putting a title's objects on one link (`strict-priority` or `fifo`; `datagram` or
  `subgroup` carriage with delivery timeouts that drop or reset; a layer cap; a VOD
  window or live capture) and a sink decoding and presenting them (workers, `base-first`
  or `fifo` decode, playout deadlines). Its figures are labelled simulated. Bases are
  sent, and decoded under `base-first`, in the order they became available. A
  refinement's stall of its base is then bounded by one chunk (strict priority) or a
  timeout and one chunk (`fifo`), plus one refinement decode (`base-first`): §10.4
  proves it, for every plan with no base timeout, a bounded link schedule and
  `base-first` decode. It is also checked on generated plans drawn over that whole
  range and on the corpus's plans, and for each unbounded part a counter-plan's stall
  passes the bound of the same plan with that part restored. The corpus `sim/` and an independent
  `sim_verify.py` reproduce every result.
- **Wire.** No envelope, payload, receipt, claim, pin or corpus file of an earlier
  revision changes bytes.

### Delivery telemetry: CMCD version 2 and CMSD for WAI objects

- **The extension** [delivery-telemetry](extensions/delivery-telemetry.md): an unsigned,
  advisory observation about WAI delivery (no joule figure, no acquisition class, no
  quality figure; content and receipt hashes only as pointers), read strictly with
  exact members; its CMCD version 2 report, in Request mode for requests and Event mode
  otherwise (`e=rr` for responses, play states, custom events for switches, dispatch
  outcomes, layer-cap changes and dropped frames), with header shards; custom keys
  under `science.transaction.wai-`, each carrying a String; switch reasons; CMSD-Static
  and CMSD-Dynamic for a WAI object; session metrics from one session's log (video
  start time, exit before start, start failure, stalls, engagement, units, counted
  bytes); and privacy rules: a content id is a content hash or a short token, and an
  emitter puts no address, query, token or personal data in it; an object URL holding a
  `?`, `#`, `@` or `\` is refused rather than stripped, so no query, signed-URL token or
  user information enters an observation however a parser splits the URL; and a
  Request-mode report goes only to the host serving the object. Cohort assignment and capable share are not in this revision.
- **Reference implementation.** `wai-rs` feature `telemetry` (no dependency;
  wasm32-clean): `telemetry::{sf, observation, cmcd, cmsd, qoe}`, with the RFC 9651
  subset the reports use; the `wai_telemetry` tool.
- **The corpus** [`telemetry-conformance/`](telemetry-conformance/): observations of
  every event kind with their reports, refused observations, CMSD descriptions and
  session logs. `wai_telemetry verify` requires every value; a differential encodes
  every report with an independent open-source CMCD and CMSD library at locked versions
  and requires the same bytes, reads them back, and validates them; CI runs both.
- **Other text.** [moq-streaming-format](extensions/moq-streaming-format.md) §5 requires
  that telemetry **MUST NOT** restate the per-object property's figure, class or evidence;
  [energy-exposure-mapping](research/energy-exposure-mapping.md) gains a dated
  addendum.
- **Wire.** No envelope, payload, receipt, claim or corpus file of an earlier revision
  changes bytes.

### The integer codecs' payloads pinned in place: `WIV1`, `WIH1`, `WIA1`, `WIS1`

- **Pinned in place** (SPEC §5, *Payloads of the WAI-native integer codecs*, and the
  new extension [integer-payloads](extensions/integer-payloads.md)): the payload of
  `wai.video.int_motion` is `WIV1` with keyframe 0; of `wai.video.int_hyper`, `WIV1`
  with keyframe 1, a nested `WIH1`; of `wai.neural.int_hyper`, `WIH1`; of
  `wai.audio.int_codec`, `WIA1`, which carries the sample rate; of
  `wai.splat.int_codec`, `WIS1`, which carries each attribute's presentation range.
  These strings' rows named a decoder and no byte layout; three of them
  (`wai.video.int_motion`, `wai.audio.int_codec`, `wai.splat.int_codec`) also named
  the tool that wrote a JSON bitstream, which SPEC §5 called the bytes the codec
  normally emits. So the pin changes what these strings carry. SPEC §8 gains
  *Completing a registration*, which permits it for a WAI-native string whose row
  fixed no byte layout and lists the envelopes it makes non-conforming; from this
  revision their formats are fixed as SPEC §8 states. No new string is registered.
  `codecs::PINNED_INT_PAYLOADS` is the one table of strings and magics.
- **The JSON is a converter input, never a payload.** The encode and export tools
  still write JSON; `int_payload::{wiv1,wih1,wia1,wis1}_from_json` convert it and
  state what it left implicit (an absent `mv_precision` is 1; the first frame carries
  no motion vectors; a short motion-vector list is padded with `(0, 0)`; the audio
  sample rate comes from the export's `audio.json`; a splat range must be exactly a
  binary32 value; the JSON `afrac` must be the model's; the channel count must match
  the medians, tables and offsets). *Envelopes this makes
  non-conforming:* a `WAI1` or `WAI2` that carries anything but the pinned format
  under these strings, a tool's JSON output among them. It still parses; a sink
  refuses it at decode (a JSON form with `bad_magic`) and does not try the fallback.
  Every emit path refuses to write it: `Wai::to_emit_bytes` and
  `WaiMulti::to_emit_bytes` (so the C packers `wai_envelope_pack` and
  `wai_envelope_pack_multi`, `WAI_ERR_INVALID_ARG`), `wai-web`'s `pack` and
  `packMulti`, and `wai wrap`. They run `container::check_int_payload_framing`, which
  parses the payload with the wire parser a sink runs and the capability's keyframe
  kind, so it checks every parse rule of the extension's §2–§6 and leaves to the
  decode only what needs a parameter set or the streams (§7); `wai inspect` reports
  what it finds. A refused emit is the new `EmitError::IntPayload`. An encrypted
  envelope's payload is ciphertext (SPEC §6), so the emit paths skip it and a sink
  checks the plaintext after decrypting (SPEC §6.4); `wai inspect` says so. No
  committed envelope or conformance corpus carried such a payload. The reference
  dispatchers (`Wai::dispatch_verified`, `WaiMulti::pick`, and so the neural sink,
  its C ABI and `wai-web`'s `dispatchCapability`) implement SPEC §6.4 step 2: they
  open no scheme, so an envelope whose manifest declares an encryption other than
  `"none"` is inert at them (`unsupported_encryption`; `container::Inert::Encrypted`,
  `container::Skip::Encrypted`, `neural::DecodeError::Encrypted`), and a payload sent
  in the clear under a claim of encryption is not decoded as plaintext. A `WAI2`
  manifest's `security` member (now stated in SPEC §7.1) must be a security object to
  be written (`EmitError::Security`, `container::manifest_v2_security`); one that is
  not is read as encrypted.
- **One integer decode path.** `int_payload::decode_int` decodes the strings above
  and `wai.neural.int_synth` and `wai.neural.int_mlicpp` from parameter files held in
  memory, with no ONNX runtime and no filesystem, natively and for `wasm32`, through a
  dispatchable capability only. A refusal is one code from a closed list
  ([integer-payloads](extensions/integer-payloads.md) §8). `int_payload::resolve`
  verifies a pin over sets held in memory and runs SPEC §3.1's check 5 for a pinned
  decode and over the set an unpinned decode reads. The neural sink delegates every
  integer capability to it: integer video, PCM and splat attributes come back as
  `neural::Decoded::Int`, never converted; `ModelRegistry::builtin` advertises
  `wai.video.int_motion`, which reads no set; the C ABI refuses an `Int` output. A
  capability this build has no integer decoder for (`unsupported_capability`) moves
  dispatch on to the fallback, as SPEC §4 treats a capability a sink cannot decode.
  Check 5 covers every identity of the bytes a decode reads (SPEC §3.1): after a pin
  verifies, it runs again over the listing of the members read and a one-file set's
  file digest (`int_payload::read_identities`), so a set revoked under one identity
  is not reached through a pin naming it by another; and a listing beside an
  unpinned set counts only when it matches the files read. Outside the integer path,
  at a sink that holds a key log, an unpinned decode reads what it can read once —
  the registered file of a one-file set, or every file below the directory its
  decoder reads from (`pin::read_dir_set`) — and check 5 runs over every digest a pin
  over it could name: the file's, or the directory's derived listing
  (`pin::listing_of_dir`) and each file's, and each listing beside them that matches.
  The decode reads only those bytes, never the path again; where the key log holds
  the capability's sets to jwp-receipts §7.9, only the part §7.9 authorised (the
  whole, the files a granted listing names, or one granted file). So a revoked set
  is refused with no listing beside it, a granted listing authorises the files it
  names and nothing else, a file changed after the check is not read, and a file the
  decode needs outside the set is `set_incomplete` (SPEC §4, *Module authorisation*).
  `pin::PinAuthority` gains `requires_grant`, with a default.
  *Breaking:* `neural::Decoded` gains the variant `Int` and is now `#[non_exhaustive]`,
  so a `match` on it outside this crate needs a wildcard arm; `neural::ParamSource`
  gains `UnpinnedRead` (an unpinned integer set read into memory once) and
  `Checked` (what an unpinned decode at a sink with a key log reads, read once) and is
  `#[non_exhaustive]` too; `int_payload::IntParams` is opaque, built by
  `int_payload::resolve` (or `IntParams::none()`).
- **Checked decoders.** `int_video`, `int_audio` and `int_splat` decode a typed
  stream that the binary and JSON forms both read into. Every size is capped at
  `2^26` elements (a `WIV1` clip at `2^30` bytes and `2^20` frames, its motion-vector
  field at `2^26` elements) the same on every target, every network's shapes are
  planned before anything is allocated, every latent and layer is checked for
  exactness, and a stream that runs out is refused at the first symbol it cannot
  give, so a build with overflow checks and one without accept the same bytes. A
  frame, a clip or a latent grows only as its symbols are decoded, to its declared
  size and no further, so a short stream stops early; `int_video` holds frames as
  RGB8. A stream that carries its symbols is decoded in full, and a few kilobytes can
  code a clip at the cap, so the header bounds the work: a sink reads the declared
  cost before decoding (`int_payload::declared_cost`, `container::declared_size`, the
  C ABI's `wai_int_payload_declared`, `wai-web`'s `intPayloadCost`) and applies its own
  size policy (integer-payloads §7, *Memory and work*). The cost counts the networks'
  multiply-accumulates layer by layer, exactly (`DeclaredCost::macs`, a `u128`;
  `macs` on `intPayloadCost`, a `BigInt`). The reference sink applies one:
  `neural::ModelRegistry::with_cost_limit` refuses a payload whose declared cost
  exceeds the limit before any symbol is decoded, with the new code `cost_over_limit`
  (`int_payload::check_cost`, `neural::DecodeError::OverCostLimit`), and dispatch moves
  on as for a capability the sink cannot decode (§8). `neural::declared_cost` reads the
  cost over the set a registry resolved, pinned or not. The C ABI's
  `wai_int_payload_declared` takes the set's path (or null for the header alone) and
  gives the working bytes and the multiply-accumulates too (`wai_int_declared_t`
  gains `working_bytes`, `macs_lo`, `macs_hi` and `cost_known`); its neural decode
  entry points no longer decode integer video, PCM or splat attributes, which they
  have no output kind for, and refuse them before decoding. A `WIA1`'s latent may be no
  longer than its `t_out` needs (`geometry_mismatch`). The JSON reader
  (`int_video::VideoJson`) caps every size before it pads a motion-vector list, and
  reads a field that does not fit the target's `usize` as absent rather than
  truncating it. In the browser, `decodeWiv1` keeps the decoder's frames
  (`DecodedVideo.frame(i)`) instead of copying the clip into one buffer.
- **One refusal order.** The extension states the order of a decode's checks
  exactly (§7): the set's members, the set against the payload, every network
  planned with its caps, the shapes, then each stream symbol by symbol, a run-out
  before the value it would give. Networks are planned layer by layer in 64-bit
  arithmetic (`int_hyper_synth::plan`, the audio synthesis's plan), a size past 64
  bits being past the cap, so a 32-bit and a 64-bit sink refuse a malformed network
  with the same code; a planned network's parameter lengths are checked before any
  symbol is decoded, and a convolution's padded side is capped. A set's entropy
  tables obey §2's length rule (3 to 4097 entries). The parser reads a table of one
  per channel table by table. A `WIV1` keyframe's planned size is checked against the
  clip's `h × w` before any of the keyframe's symbols is decoded, so
  `int_payload::declared_cost` and the decode refuse the same bytes with the same
  code; the Rust verifier checks that for every fixture. Fixtures pin each case.
- **One header for a one-frame clip.** A `WIV1` of one frame codes no motion
  vector, so its `mv_precision` and `block` must be 1 (`field_range` otherwise); the
  converters write them so. The extension's §2 says a payload's hash identifies the
  payload, not its output (unused table entries, say, code the same clip in other
  bytes), and that one output is identified by its decoded digest (a clip's
  `clip_hash`).
- **Canonical streams.** After its last symbol, each rANS stream of these payloads
  must have been read to its last word and end in the state `2^31`, where an encoder
  starts (`rans::RansDecoder::is_canonical_end`); otherwise the payload is refused
  with the new code `stream_noncanonical`. An unread word or another starting state
  coded the same symbols in other bytes. Every committed stream, the JSON forms'
  included, ends canonically. `int_hyper_synth` gains
  `decode_conditional_checked`; `int_synth` refuses a side of `2^64` or more instead of
  shifting by it.
- **Parameter sets.** SPEC §5 names the members the integer decoders read:
  `g_s.whm`, `h_s.whm` and `entropy.json` for `wai.neural.int_hyper` and
  `wai.video.int_hyper`. `wai.audio.int_codec` (`model.wam`) and
  `wai.splat.int_codec` (`g_s.whm`) are registered as one file, which widens their
  registrations (SPEC §8). `wai registry export` writes nothing new: no exported field
  records a set's shape. A `WHM1` or `WAM1` network holds nothing after its last layer
  (`IntModel::from_bytes` and `AudioModel::from_bytes` refuse trailing bytes; every
  committed network has none), so one network has one file and one digest.
- **Receipts.** A `wai.video.int_motion` receipt identifies its `WIV1` payload
  (`VideoReceipt::seal_decode_wiv1`, `verify_reconstruction_wiv1`), and
  `wai_video_meter` seals over it. The entry points that read the JSON form are
  deprecated; on a conforming JSON bitstream their output is unchanged, so every
  receipt sealed over a committed one, and every pinned digest of one, still
  verifies. They now read the JSON through `int_video::VideoJson` and the checked
  decoder, so they refuse JSON the old decoder took: a stream that runs out (it
  decoded on from the exhausted state) or does not end canonically, a table that
  breaks the table rules, `q` outside 1..=65535, an `mv_precision` other than 1 or 2,
  other than one motion-vector list per frame, vectors on the first frame or more
  vectors than blocks, a vector outside `i16`, a zero size, and a size past the caps. jwp-receipts §2.1's object content hash moves to its own
  module and feature (`jwp_hash`), which `receipt_claims` enables;
  `claim_receipt::object_content_hash` re-exports it, and no corpus changed.
- **JSON readers that refuse.** `IntModel::try_from_json` and
  `CondBitstream::try_from_value` read the conditional codec's debug JSON without
  panicking; `wai-web`'s `decodeIntHyper` and `decodeIntHyperBin` read it through
  them, `CondEntropy::try_from_value` and `decode_conditional_checked`, so an empty
  model or bitstream, or sizes past the caps, are refused. `int_video::VideoJson`
  refuses a frame whose motion vectors are not a list.
- **Browser.** `wai-web` decodes the four formats (`decodeWiv1`, `decodeWih1`,
  `decodeWia1`, `decodeWis1`); the entry points that read the JSON forms are for
  debugging and conversion.
- **Conformance.** The payloads sit beside their JSON sources in
  `wai-rs/tests/vectors/`, written by `wai_int_payload_vectors`, with refusal fixtures
  and an index in `int-payload-conformance/`; `int_payload_verify.py` checks them
  independently in Python's standard library and decodes the `WIV1` keyframe-0
  payloads in full; `byte-exact-conformance` parses and decodes the `WIV1` and `WIH1`
  payloads on every leg of its matrix. CI runs the new tests in both build profiles,
  the generator in the corpus loop, the verifier, a `cargo check` of each integer
  feature alone, and a `wasm32` check.

### The registry as one table, candidate capabilities, the codec glossary under test, and the public text retrofit

- **Candidate capabilities** (SPEC §5): a registered string whose payload format no
  revision has pinned — `wai.tensor.nnc` and `wai.splat.sog` — MUST NOT be emitted, and a
  sink treats it as unsupported. This is a MINOR rule: it enforces the candidate
  registrations as they were made (§8). Envelopes it makes non-conforming: a `WAI1` naming
  a candidate as its `capability` or `fallback`, and a `WAI2` with a component of a
  candidate; they still parse. `codecs::is_candidate`, `Wai::check_candidates` and
  `WaiMulti::check_candidates` report them. `Wai::to_emit_bytes` and
  `WaiMulti::to_emit_bytes` check, then serialise: the C packers `wai_envelope_pack` and
  `wai_envelope_pack_multi` emit through them, the browser writers `pack` and `packMulti`
  refuse the same envelopes (`container::check_candidates_v1`,
  `check_component_candidates`), and `wai wrap` refuses a candidate before any other rule,
  in every place including a `WAI2` companion. A candidate's licence class is provisional:
  `wai.splat.sog` is classified patent-pool, conservatively (it was royalty-free, never
  established). `wai.splat.sog` no longer claims an MPEG attribution: no MPEG page or
  output document read names the format it was registered for ([research
  registry](research/media-codec-registry.md), *MPEG's Gaussian-splat and point-cloud
  work*).
- **One registry table** (`wai-rs/src/codecs/registry.rs`, no feature): every registered
  string with its **kind** (`payload`, `companion`, `derivation`, `record`), its
  **status** (`registered` or `candidate`; `RegistrationStatus`), its **media** class,
  the SPEC §5 heading that registers it, the edition its registration names and the date
  it was registered; the named identifiers that are not capabilities (`wai.det.fixed64`,
  `wai.world.sim.*`); and the receipt labels no revision registers
  (`wai.quantum.{mitigate,decode,compile,rearrange,qbom}`). SPEC §5 states the kinds:
  records and derivations name no payload format. Every payload and companion
  capability carries a licence class (records do not).
- **Dispatchable** (SPEC §4, normative): steps 3 and 5 take only a dispatchable
  capability, and a capability is dispatchable unless this revision registers it as a
  candidate, a companion-only capability, a record or a derivation. A string this
  revision does not register stays dispatchable — a later MINOR revision may register
  it (§8) — and is decoded only where the sink advertises it. `codecs::is_dispatchable`
  implements it and is the one predicate every dispatch and selection function of the
  reference implementation uses (`Wai::dispatch_verified`, formerly
  `dispatch_capability`; `WaiMulti::pick`; `WaiMulti::primary`;
  `RenditionMeta::is_selectable`; the CLI's `inspect`; and, below, the neural sink, the
  C ABI and wai-web); §7.1's selection sentence says the same. A later dispatch or
  selection function uses it too.
- **`wai registry export`** writes `registry/capabilities.json`, its JSON Schema
  (`registry/capabilities.schema.json`), the generated table of
  [`REGISTRY.md`](REGISTRY.md) and **vocab v2** (`vocab/v2.{ttl,jsonld,shapes.ttl}`,
  namespace `https://wai.transaction.science/vocab/v2#`, one IRI per registered string,
  enumeration individuals named `<class prefix>-<label>` so no two terms share an IRI,
  the registration date as `registeredOn`; its JSON-LD context maps a `WAI1` or `WAI2`
  manifest as written, with `kind` and a component's `role` under terms of their own;
  its SHACL shapes apply to a manifest with no `@type` and to every `WAI2` component,
  check the capabilities against the registry, and leave `media` open, as SPEC §3 and §8
  require). `--check` fails on drift; a library test and a CI step run it, and
  `vocab/check_shapes.py` validates typed and untyped sample manifests in CI. vocab v1
  is stable: no IRI is removed and no term changes meaning. REGISTRY.md states the
  registration procedure; CONTRIBUTING.md is rewritten around it.
- **The `wai` binary builds without the `codecs` feature** (encode and decode are
  compiled out); `tests/cli_wrap_intent.rs` and the new `tests/cli_registry.rs` run in
  CI. `wai wrap` now writes the registry's media class: `volumetric` for `wai.mesh.*`,
  `wai.pc.*` and `wai.splat.*` (was `model`), `image` for the learned image codecs such
  as `wai.neural.bmshj2018` (was `binary`). SPEC §3 lists the media values, informational.
  Without `codecs`, `inspect`, `capabilities` and `unwrap` claim no native decode.
- **Editions** (SPEC §8): a registration that names no edition carries the edition of
  its standard current on the date it was registered, and no later edition or
  amendment, which takes a new string. `wai.pc.gpcc` and `wai.pc.vpcc` therefore do not
  carry the Gaussian-splat amendments of ISO/IEC 23090-9 and 23090-5; Enhanced G-PCC
  (ISO/IEC 23090-38) is not a `wai.pc.gpcc` payload. Which edition of each part was
  current on their registration date is unresolved, and what they admit beyond the
  part's first edition is unspecified until a revision records it.
- SPEC: `wai.world.replay`, `wai.world.live`, `wai.semantic.clip` and
  `wai.semantic.perco2` are spelled out or given rows, and a test holds SPEC and the
  registry to the same strings in both directions; the opening, §1 and the floor
  paragraph agree with §5; citations: PNG (W3C Recommendation, Third Edition), FLAC
  (RFC 9639), Zstandard (RFC 8878), AVIF (the AOM AVIF specification, HEIF-based,
  MIAF-conforming), JPEG AI (ISO/IEC 6048-1:2025); the neural tables state each model's
  published operating point; Appendices B and C are rewritten; confidentiality's
  implementation is a companion implementation under its own licence.
- **The codec glossary** links each registered capability that has an entry, with the
  registry's licence and replicate classes, through a reviewed table
  (`GLOSSARY_LINKS`); every other payload and companion capability is listed with a
  reason (`GLOSSARY_NO_ENTRY`). `org` became `body`, from a closed allowlist of
  standards bodies, designations and open-source projects; new modality values `text`,
  `volumetric`, `haptic`, `metadata`; new entries for the registered formats that had
  none; descriptions scrubbed of company names and comparisons; entries titled by the
  format's own name, FourCC, file extension or designation, never by a company, a
  product or a person (the research codecs by their method); HEVC regraded `success`;
  JPEG AI dated 2025; AVS3 video's body is the designation the registry pins. The file
  now carries `"schema_version": 2`: a reader of the earlier form, which had `org` and no
  version, finds no `org` field in version 2.
  `glossary_links_agree_with_the_registry` holds it to its schema and the registry.
- **quantum-ops vectors regenerated with neutral ids** (`example:lab3-device`,
  `example:qpu-a`) through `wai_quantum_ops_vectors gen`. Changed receipt hashes:
  `cal_q3` `7c53298929a1261959315c91f9540781b4bff0669b5f4b783becee93dbf5f8fa`, `cal_cz`
  `4c6374ee05c956f928455f86326357f10d08f3466e88d569a6ef49975001f526`, `job_bell`
  `08b142d29dadc466b294b89d2127f39ee9fa1aa0b6b5a4c6c5d226baf343b9e3`. The wai-quantum
  calibration pin (`receipt_pins::calibration_legacy_bytes`) was recomputed by the code
  before the acquisition class, with the neutral id. wai-quantum's version is unchanged.
- **CI**: the feature-gated test targets the conformance workflow names are declared
  with `required-features`, so dropping a feature stops cargo instead of running nothing;
  `tests/int_audio.rs` and `tests/int_splat.rs` now run; the registry export is checked;
  the glossary, SPEC, registry and vocab paths fire the conformance workflow. wai-web is
  built for wasm32 (`wai-web-wasm.yml`, and the int_hyper vectors in
  `wai-sota-vectors.yml`), and its verify scripts fail on a missing artifact when
  `WAI_REQUIRE_SOTA_VECTORS` is set or a directory is named. The Windows x86_64 legs are
  out of the conformance and byte-equality matrices while their runner is offline (each
  matrix says how to restore it). The nested `wai/.github/workflows/ci.yml`, which no
  forge reads, is removed.
- **Public text**: two checks, `scripts/check-comparative-prose.sh` (a public phrase
  list) and `scripts/check-public-names.sh` (patterns kept private), now run in CI over
  `wai/`'s text, code, data, markup and ignore files. Both fail closed (a grep error or
  a pattern that does not compile exits 2), an opt-out needs a reason, and
  `scripts/test-public-text-checks.sh` proves both on every run. Without the private
  list a push to main fails. The list is given only to the run on a push to main, never to
  a pull request's code; a pull request shows a failed step saying the name check did not
  run. The specification, extensions, research notes, READMEs, CONTRIBUTING, the
  glossary and source comments name no company, vendor or product and make no
  comparisons; functional identifiers (capability strings, model-hub ids, OS API names,
  advisory ids) stay, marked. Dated research records keep their findings: corrections are
  dated annotations (the MoQ Transport revision, the MPEG splat stages, WebTransport's
  status, the October 2026 call for content); names are redacted, and comparative
  framing is restated, in place. In this changelog's history, names are redacted and
  nothing else changes.
- For later changes: a new dispatch or selection path uses `codecs::is_dispatchable`;
  a newly registered string (for example a prior set) goes into the registry table,
  `wai registry export` regenerates the exports and vocab v2 (vocab v1 is stable and only
  gains terms), and the string is listed in the glossary's `GLOSSARY_NO_ENTRY` or linked.
  The quantum-ops corpus and the wai-quantum calibration pin have the values above; the
  pwsh steps of the conformance workflow remain, inactive.
- [`extensions/jwp-receipts.md`](extensions/jwp-receipts.md) §5: the MUST is over the
  cases §5.1–§5.5 define; verifying against their vectors becomes a MUST once the
  vectors are published in the tree.

### The parameter-set pin, specified; component identity; pin-aware selection

- **SPEC §3.1** defines the pin a manifest carries (`conditioning.prior`; a `WAI2`
  rendition's `prior`): `id`, `sha256`, `determinism`, and the optional `form`,
  `tier` and `supersedes`. The digest covers the whole parameter set: a form-less
  pin is admissible only for a capability registered with a one-file set (§5
  registers each capability's shape), and a set of several files is pinned through
  its listing, the new registered prior form `wai.prior.set` (`StandardDefined`,
  royalty-free, the one form listed as established exact). A sink verifies the pin
  before it decodes, reports a refusal with one of an ordered set of snake_case
  reasons (`pin_malformed`, `form_not_registered`, `form_unreadable`,
  `form_required`, `prior_revoked` / `prior_not_authorised` (check 5, run by a sink
  that holds a key log; see *Key discovery, signing authority and module revocation*
  below), `prior_not_held` / `digest_superseded`, `set_malformed`, `part_mismatch`;
  `set_incomplete` during a decode), and decodes only from the set it verified. A pin over a capability this
  revision does not register is verified as several files. `null` in an optional
  member is the same as absent.
- **Tier** is a closed vocabulary (`decode-equivalence`, `entropy-consistency`,
  `none`): an encoder writes a registered token, a sink reads any other as `none`,
  a verifier reports it (`tier_unregistered`). Pins without `tier` keep the legacy
  reading of `determinism`, now stated normatively, so every existing verdict —
  `tests/vectors/int_mlicpp/stream.wai`'s included — is unchanged.
- **`supersedes` is informative**: it changes only the reason a sink that holds the
  predecessor reports, a malformed one is a static finding the sink ignores, and a
  successor never satisfies its predecessor's pin.
- **§4 and §7.1**: a pin that does not resolve makes its capability unsupported for
  that envelope; dispatch continues at the fallback or the next rendition, and an
  inert envelope's error names the reasons: for `WAI1`, the reason kept for the
  capability before the fallback's, `set_incomplete` included; for `WAI2`, every
  entry's (`not_selectable`, `missing_capability`, `component_digest_mismatch`, a
  pin reason, or, under a policy that serves only sound `replicate` claims, the
  entry's `replicate_*` reason). A `WAI1` pin applies to whichever capability
  decodes the payload.
- **Component identity (§7.1)**: a rendition entry MAY carry `id` (unique, never a
  capability string), `sha256` (its payload digest, checked before the entry is
  decoded or applied: `component_digest_mismatch`) and, on a companion,
  `companion_of_id`, which binds it to one entry — an enhancement layer can now
  bind to one of two renditions of a base capability. `id`, `sha256` and `prior`
  are one construct; no extension defines another per-component digest.
- **§7**: `replicate` over a `NeuralIntegerExact` capability MUST carry a pin that
  declares `decode-equivalence`, in no form or in `wai.prior.set`. The new
  *Digests* subsection names SHA-256 for parameter-set identity, for component
  identity, and for the decoded output of capabilities that decode against a
  parameter set and of standard codecs in conformance material; FNV-1a and BLAKE3
  stay as diagnostics. The vectors of such capabilities gain `sha256.json`
  sidecars; no existing vector file or value changes.
- **§8** settles how an optional member is added under the 1.x freeze, which the
  `form` entry below left open: a MINOR revision MAY define an OPTIONAL member,
  envelopes keep `wai` `"1.0"`/`"1.1"`, a reader MUST ignore a member it does not
  understand whatever the MINOR, and an earlier reader may instead reject such an
  envelope (fails closed) or ignore the pin (fails open), which §3.1 states. §8
  states separately the one exception this revision makes to the freeze (next
  entry).
- **Envelopes made non-conforming:** the list in SPEC §3.1. None is committed.
  One kind stops parsing, the one exception to the `1.x` freeze that §8 states: a
  manifest that carries a member this revision defines (a pin or pin member, or a
  rendition entry's `id`, `sha256` or `companion_of_id`) with another JSON type is
  malformed, because reading a mistyped `form` as absent would turn a set-form pin
  into a form-less one. The envelope's `wai` value does not signal it.
  New vectors: `tests/vectors/int_mlicpp/{model.sha256, stream_set.wai}` (the
  set-form twin, written by `tools/wai_int_mlicpp_set_form.py`) and the
  `pin-conformance/` corpus, with an independent verifier in Python.
- **Reference-sink behaviour:** a form-less pin over a several-file capability is
  refused (`form_required`), where before it verified one file or failed on a
  directory; `wai.neural.int_synth`'s pin covers `int_synth.json`; a digest that is
  not 64 lowercase hex is refused, where before it was compared case-insensitively;
  a refused pin falls through to a fallback the registry holds; every decoder reads
  parameters through `neural::ParamSource`, so a pinned decode reads only verified
  bytes (ONNX sessions built from them), and an ONNX model that keeps a tensor's
  data in another file, which the runtime would read from outside those bytes, is
  refused with `set_incomplete` naming that file, as §3.1 now states; `bmshj2018`'s CDF cache is keyed by content
  digest and holds a bounded number of tables, least recently used evicted.
  `decode_envelope_strict` applies §7's rule to the capability it decodes through,
  so a sound fallback behind an unsound primary the sink cannot use still decodes;
  it refuses before reading a file only when every capability it could decode
  through (dispatchable and advertised) is refused, so a candidate the registry
  happens to hold does not change its reason. A `WAI2` parse sums the rendition lengths in 64 bits, so a
  table whose lengths wrap a 32-bit sum is refused as truncated on 32-bit targets
  too, where it used to panic. `wai inspect` prints digests through
  `pin::digest_preview`, which shows a malformed one whole instead of cutting it
  inside a character.
- **Rust API, source-breaking** (the same 0.4.0 release as `form`): the manifest
  types (`Manifest`, `ModelRequirement`, `Conditioning`, `Prior`, `Supersedes`,
  `ManifestV2`, `RenditionMeta`, `Rendition`, `Security`) are now
  `#[non_exhaustive]`, with constructors and `with_*` builders that write the same
  JSON a literal wrote, so the next optional member does not break callers again;
  every manifest object keeps members it does not model (`extra`), and `Manifest`
  gains `security`, so a Rust round trip loses nothing. `Rendition` holds a
  `RenditionMeta`. `Wai::dispatch_capability` is replaced by `dispatch_verified`;
  `WaiMulti::pick`/`primary` take a pin verifier and return `Picked` (index, entry,
  payload, verified set, skipped entries); `companions_of_index` binds by id.
  `DecodeError::PriorMismatch` carries a `PinRefusal`, and `DecodeError` gains
  `NoRendition`; `ModelRegistry` gains `held`/`hold`/`verify`/`resolve_verified`
  (`resolve_checked` is deprecated); new `decode_payload`, `decode_multi`,
  `decode_multi_strict`. `sha2` is now a plain dependency, and the new `pin` module
  needs no feature.
- **C ABI:** a codec-free `ffi` feature carries the envelope surface (`codecs`
  implies it); `WAI_ERR_PRIOR` (5), `WAI_ERR_UNSOUND_REPLICATE` (6) and
  `WAI_ERR_NO_RENDITION` (7), with the reason first in `wai_last_error`; the unpack
  calls return the original manifest bytes, as `wai.h` already documented; the
  `WAI2` table carries identity, pin and `selectable`; new `wai_envelope_select`
  and `wai_neural_decode_ex` (`WAI1` or `WAI2`, a registry of held sets, a strict
  flag), which writes `*out` on every call, empty on failure, so `out->data` is
  always safe to free. The envelope packers build under `ffi` too, and
  `tests/ffi_envelope.rs` checks in CI that they refuse a candidate and that
  `wai_envelope_select` never picks one.
- **wai-web** wraps `wai::container` instead of reduced copies, so `role`,
  `companion_of`, identity and pins reach JS, and exposes `pick`, `companionsOf`,
  `verifyPin` and `sha256Hex`; a component whose capability is not dispatchable
  (SPEC §4) is never presented as media, and `rendition(i)` and `companion(i)` refuse an entry whose payload does not
  match its `sha256`. Its JS-facing tests run under wasm32 in CI.
- **One dispatch predicate:** `dispatch_verified`, `pick`, `primary`,
  `RenditionMeta::is_selectable`, `decode_multi`, `decode_envelope_strict`'s early
  refusal, `wai_envelope_select`, wai-web and `wai inspect` all take a capability only when `codecs::is_dispatchable`
  (above) accepts it, so a candidate, a record or a derivation is never dispatched
  or picked, whatever its pin. An unregistered string stays dispatchable where the
  sink advertises it, and a pin on it is verified as several files, so it fails
  closed. SPEC §7.1 step 2 reads "dispatchable (§4)". `wai.prior.set` has its row in
  the registry table, so `registry/capabilities.json`, `REGISTRY.md` and vocab v2
  list it.
- **Vocabulary:** `vocab/v1` gains the pin terms additively, and `wai registry
  export` writes the same terms and shapes into v2, with the tier list taken from
  `codecs::TIER_TOKENS`; `vocab/check_shapes.py` validates conforming and
  non-conforming pins against the v2 shapes. `sharedPriorUri` and
  `sharedPriorHash` are annotated as not being the manifest pin. The SHACL pin shape
  targets the objects of `wai:prior` (no class is asserted for them), and
  `supersedes` is checked by its own shape: a required digest and an optional form.
- **Open item:** `wai.world.model` names its renderer's model by `model_id`, a
  second model-identity construct; it is to fold into the pin once that renderer
  is registered as decoding against a parameter set.

### Energy reports, labelled object leaves, and what the signatures cover

- **Reports about a figure.** Two claim classes, [`extensions/jwp-receipts.md`](extensions/jwp-receipts.md)
  §2.4.8–§2.4.9, report on one energy figure without changing the receipt that seals
  it. A claim names that receipt by a **figure profile** — the receipt's identity domain
  less its version byte, so the registry is the code's own
  ([`extensions/energy-measurement.md`](extensions/energy-measurement.md) §6.1) — and its
  identity, and restates the figure with its class. `measurement` carries the bracket's
  values as integers with each bracket's liveness facts, the build (profile, toolchain,
  target) and the power reader; only the key that signed the figure may sign it, so a
  figure no Ed25519 key signs (an energy binding's) has no `measurement` claim. A
  hardware figure's report must name its power reader (`measurement_no_power_reader`),
  with or without a bracket, and a `null` bracket says the figure was not obtained under
  the bracketed protocol. Because only §3.1 step 6 shows the signer is the figure's, a
  `measurement` claim is accepted only against the receipt that carries its figure:
  checked without it, it is `figure_not_held`, and `ClaimReceipt::attribution` gives a
  verdict only for a claim accepted against that figure.
  `operating-point` records the requested, committed and delivered operating points,
  from a vocabulary named by URI, and a prediction. energy-measurement §6 makes the
  `measurement` claim the carrier of a class-bearing figure's report and lists the
  producers that fall short.
- **Exact report arithmetic.** energy-measurement §6.2 gives the figure and the least
  declaration a report's integers support as exact integer formulas, at whatever width
  their products need. Producers seal from the same integers; computed from unrounded
  samples, a declaration could come out one ppm below what the published integers
  support. `wai-rs` `energy_report` implements it with no floating-point arithmetic, and
  `claim-conformance/claim_verify.py` recomputes the pinned brackets with Python's
  integers.
- **Derived, never stated.** The §4 ratio and verdict, whether a commitment was met,
  and a prediction's residual are derived by every reader from what a claim carries. A
  report carrying its ratio is refused (`unknown_member`).
- **A live power reader.** energy-measurement §3.1: a rail is live in a bracket only if
  it advanced in two consecutive intervals, in at least five and in at least four of
  every five, over at least 300 ms, in both brackets; a figure from a bracket that is not
  live is never sealed under a hardware class. A power sample advances only when it is
  above zero and differs from the sample before it, so a reader that holds its last
  value never advances; a reader of a report refuses, under a hardware class, a bracket
  whose `sd_uw` is 0 whatever liveness facts it carries. A frozen rail, a stale value
  held, one batch refresh, and one refresh published in two steps all used to give a
  flat idle bracket and an infinite ratio that sealed as `OnChipCounter`; `energy_burst`
  and `energy_meter` now refuse them, with fault-injection tests, and a counter meter reads its counter during a span
  rather than only at its ends. On a host whose reader reports a rail that does not
  advance, the reference meters seal unmetered and say why.
- **The unmetered marker.** energy-measurement §2: an unmetered figure is not a figure
  of zero joules. A `0` that marks unmetered is never presented, totalled or compared as
  0 J; a total over parts that include one covers only the measured parts; new forms
  mark unmetered with class tag `0`. quantum-sim, provenance, interactive-worlds,
  moq-streaming-format and SPEC read `0` this way. `ModelMeter` returns no reading rather
  than a labelled zero or a saturated product, and `wai-quantum-serve`'s `POST /verify`
  renders an unmetered receipt as `"joules_micro": null, "metered": false`.
- **The labelled JWP leaf.** An object's class rides in its Merkle leaf,
  `BLAKE3("moq-jwp:merkle-leaf\x02" ‖ content_hash ‖ u64_be(joules_micro) ‖ class ‖
  u32_be(len(origin)) ‖ origin)`, so the group signature covers the object's figure, its
  class with its evidence, and its origin (jwp-receipts §2.2, §2.4.2). A label, or the
  origin, can be neither added, stripped nor altered in transit.
- **The labelled group payload.** `moq-jwp:group-sign\x02` signs the legacy fields and
  then the total's class, its coverage and the parent link (§2.3, §2.4.1); the legacy
  payload never signed the parent. The class is composed from the objects' classes by an
  integer rule — the weakest class, and for a hardware class `⌈Σ F·p / Σ F⌉` ppm over the
  longest window — and a verifier holding every object checks it (§3, *The group
  total*), alongside the sum.
- **Position.** An object must sit at the position its inclusion proof walks (§3 step
  5, `proof_position_mismatch`), which closes the duplicated last leaf of an odd level
  and an object presented at another object's index — the latter could hide a weaker
  class from the group total. Group signatures are held to the strict Ed25519 rule
  claims use. §3.3 lists the delivery reasons; a group's link hash is defined over its
  signing payload and signature.
- **`WAI_RECEIPT`** ([`extensions/moq-streaming-format.md`](extensions/moq-streaming-format.md)
  §5): `0` is the unmetered marker, `group_receipt` is the group's link hash, and the
  property's class reads as signed only where it equals the object's leaf label or an
  accepted `energy` claim; otherwise it is discarded (`property_disagrees`) or read as
  unlabelled. A class read as signed names the key that signed it — the group's for a
  leaf, the claim's for a claim — so a class a relay signs into the property under its
  own key is never read as the publisher's.
- **What stays unverifiable.** energy-measurement §5 and energy-binding §3 now say it is
  the measurement — the joules and every field describing how they were obtained — not
  one number, and that a report's values can be checked for consistency, which is not
  truth. energy-binding §7.4 no longer says green metadata is registered; it is not yet.
- **Implementation.** `wai-rs` (feature `receipt_claims`): `jwp_delivery` reads and
  verifies both forms of each JWP receipt, the group total and the property;
  `energy_figure` holds the profile registry and views every figure-bearing receipt
  family, the quantum receipts through `MaybeLabelled`; the reference meters write a
  measurement report beside the receipt by default (`--no-report` to opt out;
  `--previous-report` with no file is a usage error), and `wai_prov` presents its total as
  `Partial` over the measured steps, or `UNMETERED`, never as 0 µJ. joule-code's
  `joule-moq` (0.2.0) builds and verifies both labelled forms, checks position, holds
  publisher keys and group signatures to the strict rule (a non-canonical key encoding is
  refused as a key), refuses an unsigned coverage on a legacy group, and runs every
  leaf, group and object vector of the new `jwp-conformance/` corpus.
- **Conformance and CI.** `claim-conformance/` gains vectors for every new reason, the
  arithmetic at its boundaries, liveness faults (a held value among them) and the
  labelled leaf; the earlier vectors are unchanged and pinned. `jwp-conformance/` holds the JWP corpus
  (`wai_jwp_vectors`) and `jwp_verify.py`, an independent verifier. CI now runs
  `claim_verify.py`, `differential.py` (fixed seed) and `jwp_verify.py` on the linux
  runner, and the JWP corpus on every platform of the conformance matrix. The
  conformance build includes `receipt_claims`, so `wai_meter` and `wai_world_meter`
  are built on every platform of the conformance matrix; `wai-energy-receipts.yml`
  builds those two, `wai_video_meter` and `wai_quantum_meter` on linux, linux-arm64
  and macos; no workflow builds `wai_prov`, which needs `codecs`. joule-code's CI runs
  `joule-moq` with its QUIC transport.
- **Build.** Every reference meter bin (`wai_meter`, `wai_world_meter`,
  `wai_video_meter`, `wai_quantum_meter`, `wai_prov`) now also requires the
  `receipt_claims` feature, because it writes a measurement report. Cargo skips a bin
  whose required features are off without an error, so a build that names only the
  older features (`--features video_receipt`, say) no longer produces that meter; add
  `receipt_claims`.

**Wire.** Every receipt issued in a legacy form keeps its bytes and verifies as before:
v1 groups, `\x01` leaves, every claim and corpus vector already pinned, `wai-quantum`'s
receipts (no `wai-quantum` code changes). The new forms are refused by software that
predates them: a verifier or relay refuses a labelled group at §3 step 2 and a labelled
object at §3 step 5, and a claim verifier refuses `measurement` and `operating-point` as
`unknown_class`. A group's `sign_version` other than 2 is `group_label_invalid`.
**External JWP implementations must adopt the `\x02` leaf and the labelled group payload
before verifying objects or groups that use them, and hold publisher keys to the
canonical-encoding rule**; the in-tree `joule-moq` does, and its public API changes are
listed in joule-code's CHANGELOG. `wai-quantum-serve` (0.4.0) changes the `POST /verify`
response: an unmetered receipt renders as `"joules_micro": null, "metered": false`
where it rendered its marker as the integer `0`, and every other receipt adds
`"metered": true`; a client that reads `joules_micro` as an integer must accept
`null`. New meter receipts can declare one ppm more than before, because they
seal from the published integers, and a meter now refuses a rail that is not live where
it used to seal a figure.

### Key discovery, signing authority and module revocation

- [`extensions/jwp-receipts.md`](extensions/jwp-receipts.md) §7: a signer can say
  where its keys are published, and a verifier can confirm it. A `signer_id` of the
  form `did:web:<host>[%3A<port>]#<key-id>` names a web origin, which publishes at
  `/.well-known/wai-keys` a **key log**: revisions signed by a root key that
  pre-commits to its successor, each listing the origin's keys, the authorities each
  holds (`delivery`, `energy/<acquisition class>`, `lawful-basis/declaration` or
  `/decision`, `log-inclusion`, `soft-binding`, `binding-recovery`, `measurement`,
  `operating-point`, `receipt`, `module/decoder`, `module/reconstruct`) and whether it
  is active, retired or revoked. The path is published now; its registration in the IANA Well-Known URIs
  registry is to be requested. A `did.json` on the same origin SHOULD list exactly
  the log's active keys, and the log governs.
- A verifier that has anchored the origin — pinned out of band, or fetched from the
  origin over HTTPS and recorded as such — confirms an accepted receipt, claim or
  module grant against the log, offline, and reports `confirmed`, `refused` or
  `unconfirmed` with a named reason beside the item's own verdict, never folded into
  it. An item needs every authority the item fixes: a group receipt needs `delivery`
  and `energy/<class>` for each class its objects' labelled leaves and its labelled
  total carry, another family's receipt `receipt` and `energy/<class>` for each class
  its figures carry, a claim its class's authority, and an energy authority is read from any
  `energy_provenance` a claim's body carries, whatever its class — so a `measurement`
  or `operating-point` claim needs its class's authority and the energy authority of
  the figure it restates, pinned-only for a hardware class. An Ed25519 `did:key`
  names one key, and one naming a key other than the signer's is refused; a `did:key`
  longer than any Ed25519 key's spelling is a label, refused before it is decoded. The head walk
  ignores revisions that are not held, unattached, or signed by a key their parent
  does not authorise; the owner's rotation supersedes a stolen root's revisions; an
  older log retrieved after a newer one was anchored is reported as rolled back.
- Pinned-only authorities: the hardware-class energy authorities and the
  lawful-basis authorities, and any authority registered for pre-emption, are
  confirmed only under a pinned anchor that has not moved (`anchor_not_pinned`);
  every confirmation reports whether its anchor was pinned or fetched. A pinned record
  moved forward by a later retrieval stays pinned and keeps the revision that was
  pinned, and a pre-emption authority is read from that revision only. Beside an item
  that needs a hardware-class authority, a verifier reports whether its key also
  holds `delivery` (`also_delivers`); the recommendation that it not is a SHOULD.
- Revocation refuses every item under a revoked key, whatever time the item carries.
  A retired key's untimed items stay confirmable until the key is revoked, and are
  reported as untimed.
- §7.9, decoder and reconstruct modules: a module, or a parameter set an envelope
  pins, is identified by the SHA-256 of its bytes and authorised by a signed **module
  grant** from a key holding `module/<form>`, for named capabilities and a bounded
  window. A grant is held only from one of the sink's **module origins** — an origin
  its relying party names, or the origin under which it confirmed the envelope's
  delivery — so anchoring an origin for its receipts does not make it a source of
  modules. The key log revokes modules by digest (`revoked_modules`, append-only), and
  a digest revoked in the head of any origin the sink has anchored is not held,
  whichever origin grants it. Each log declares `grant_max_s` and `status_max_age_s`
  under normative ceilings; a signed **key status** from the head's root, re-issued on
  each heartbeat and held only if it names the head or an ancestor of it, keeps a sink
  current. Eight ordered checks decide whether a sink runs a module; every failure —
  including no network, a stale status and no trusted clock — means the module is not
  held, and dispatch falls back: a sink never reports a file inert because of module
  authorisation while a fallback remains, and an encoder whose capability needs a
  module declares one. The kill latency `grant_max_s + status_max_age_s` is a
  normative parameter. The ceilings are 86400 s for `grant_max_s` and 3600 s for
  `status_max_age_s`.
- Module revocation through the parameter-set pin. [`SPEC.md`](SPEC.md) §4 folds the
  hook into its steps: a sink that runs a module obtained apart from an envelope holds
  it under §7.9, or the capability is not advertised for that envelope. SPEC §3.1's
  check 5, reserved until now, runs at a sink that holds a key log, keyed on the pin's
  `form` and `sha256`, before check 6. It has two parts. Revocation runs for every
  capability a pin applies to and needs no grant, status or clock: a set whose digest
  the head the sink reaches from any anchor it holds revokes is refused with
  `prior_revoked`, whichever check a grant would fail first. Authorisation runs where
  the relying party holds the capability's pinned sets to §7.9, its choice per
  capability: §7.9 authorises the set as a decoder module, and any other failure,
  offline and no grant included, is `prior_not_authorised`. Dispatch moves to the
  fallback as for any refused pin. An envelope that pins no set does not escape check
  5: the set an unpinned decode reads is checked under each digest a pin over it could
  name. A sink configured with no key log does not run check 5.
  `key_log::module_check` implements §7.9's checks; `key_log::SinkKeyLog` runs check 5
  in the reference dispatcher, given to `pin::verify_with` or to
  `neural::ModelRegistry::with_key_log`, which also runs it before an unpinned decode.
  Library tests in CI dispatch a revoked set to its fallback, report `prior_revoked`
  when none remains, refuse a revoked set whose grant fails an earlier check, whose
  sink holds no grant, or whose capability is not held to §7.9, and fall back when the
  sink is offline; the neural sink gates run the same refusals through
  `neural::decode_envelope` over the committed int_mlicpp set, pinned and unpinned. A
  key log that names the subject's origin holds for one envelope, so a registry given
  one decodes that envelope only. The reference dispatcher runs no module obtained
  apart from an envelope; `wai-web` and the C ABI take no key log.
  `neural::ModelRegistry` gains a private field, so a struct literal no longer builds
  one; build it with `ModelRegistry::new()` and its builder methods.
- `wai-rs`: `signer_id` (ungated, no dependency) reads a `signer_id`;
  `key_log` (feature `key_discovery`, no new dependency, checked to build for wasm32)
  implements §7. Outside the crate, an item to confirm is read from bytes that
  verified: `Item::from_claim` from an accepted claim, `Item::from_group` from a group
  receipt and the objects included under it, `Item::from_family` from another
  family's receipt through its figure view, given the `signer_id` and signed time the
  same receipt carries. `wai-keys` builds an origin's key log from seeds the operator names,
  appends retirements, revocations, revoked modules and rotations, signs statuses and
  module grants, verifies a log as §7.3 anchors one — afresh, or as a later retrieval
  against a fetched or pinned anchor already held — and emits and checks a `did.json`.
  It creates no key material, never prints a seed, and refuses a published test key
  unless told the log is a test log.
- `signer-conformance/`: a corpus of anchoring, confirmation and module vectors with
  declared verdicts, generated and verified by `wai_signer_vectors` in CI on every
  platform of the conformance matrix, with a canonical fixture computed by a separate
  implementation (`signer_verify.py`), which also reproduces every vector's verdict. CI
  runs it on the linux runner and requires its verdicts to equal `wai_signer_vectors`'
  line for line (jwp-receipts §5.8).
- No wire format or committed signed vector changes. Every receipt and claim verifies
  as before, and every vector of the claim corpus, those of the `measurement` and
  `operating-point` classes included, keeps its outcome. The reference meters, whose seeds are published, now write a label as
  `signer_id` rather than a real origin, so their signed output differs from before;
  the world demo's signer is a label too. A CI test holds every `did:web` literal in
  a tracked `.rs`, `.json` or `.js` file under `wai/` (`signer-conformance/` aside)
  that names a domain, and every origin-bound one, to a reserved example name. `wai-quantum-serve`'s `/verify`
  reports the signer's level beside validity.
- transparency §6 pointed to "jwp-receipts §7" when no §7 existed; it now points at
  this section. SPEC §6.2–§6.3, confidentiality §6–§7 and jwp-receipts §6.6 no longer
  say that a relay signs, or that checking a receipt needs no key: it needs the
  signer's verifying key, which the receipt carries. prior-carriage §6 no longer says
  WAI has no key discovery.
- Not done here: a key log for `wai-web` and the C ABI; an independent time that could
  reinstate an item signed before its key was compromised; sub-delegation between
  module grants; an authority for pre-emption, relay attestation or reconciliation,
  which have no class; checking a log's evidence under its profile; the keys an OER
  receipt is checked with.

### Typed receipt classes, and a second (soft) binding class

- [`extensions/jwp-receipts.md`](extensions/jwp-receipts.md) §2.4.3–§2.4.7: every
  receipt now belongs to one **class**. `group`/`object` are the `delivery` class, with
  their wire form unchanged; a new `claim` kind carries the others on one envelope (JCS,
  RFC 8785), one Ed25519 signing rule and a per-signer chain kept separate from the
  group chain. Classes: `energy`, `lawful-basis` (a declaration of purposes each with
  its own basis, or a decision recorded against one) and `log-inclusion` (the log's
  evidence carried verbatim — accepting the claim is not verifying the inclusion). An
  unknown class is rejected, never skipped; `delivery` is reserved to `group`/`object`.
- **`energy`** gives an object's figure a signed place for its acquisition class, which
  the object receipt has no member for
  ([`extensions/energy-measurement.md`](extensions/energy-measurement.md) §1 now says
  so, and that a group total still has none). The claim restates the figure and carries the
  class with its evidence in energy-measurement §2.7's JSON fields — `energy_provenance`,
  `energy_uncertainty`, `energy_counter_filtering`, `energy_calibration_ref`, present and
  `null` where a class has none — admitted through `crate::energy_class`: a refined
  hardware class needs its declared uncertainty above the resolution floor, a counter
  its filtering state and an instrument its calibration reference, and legacy `HwShunt`
  is refused (`energy_legacy_acquisition`, `energy_acquisition_evidence`). Its grade may
  not rank above what the class supports under
  [`extensions/oer-receipts.md`](extensions/oer-receipts.md) §2 — both refined
  hardware classes `measured`, `ModelBased` `derived`, `Estimator` `projected` — so a
  verifier that checks an accepted claim against the OER receipt it names holds one
  graded within the class the claim states. A named OER receipt
  is read, verified and its figure read by the OER module under oer-receipts' own rules
  (a vector document read under §4's profile, a figure read as §4 reads a quantity and
  compared exactly); the claim layer has no reading of OER of its own.
- §2.6 and §3.2: a **second binding class**. A publisher-signed `soft-binding`
  declaration names a registry algorithm (reverse-DNS form, aligned with the C2PA
  soft-binding algorithm list), its type, the value and an `https` resolution endpoint;
  a `binding-recovery` record states `attests: "ancestor"`. **The ancestor rule**: a
  receipt reached through a soft binding attests the ancestor, never the object in
  hand, and the transformation's energy is unaccounted rather than zero. A declaration
  counts only if signed by the publisher of an already-verified delivery receipt for its
  subject, and only once its own signature is verified — when a recovery is performed,
  and when a recorded one, or a lawful-basis decision, is checked against the
  declaration it names (`context_invalid` otherwise).
- §3.1 step 3 restates RFC 8032's own decoding rule (a key has one canonical encoding)
  and range rule (`S` < L), and fixes the two choices on which RFC 8032 verifiers
  differ: keys and `R` of small order are refused, and the equation is the cofactorless
  one. Step 1 states what JSON is `malformed` where parsers part — a byte-order mark, a
  repeated member name, an escape that leaves a surrogate unpaired, nesting deeper than
  32 — and that a number never is (`1e400`, `-0` and a 400-digit integer are
  `field_type`); `issued_at` must name a day its month has; `resolution_endpoint`'s host
  must be non-empty and its port digits. Step 2 checks exactly the type each member
  table gives, a member's type and encoding together in the table's order (a
  lawful-basis body's `record` first, a nested object's members where the table reaches
  it); what a class's rules add to a type — a non-zero figure or work count, a non-empty
  basis, an `alg` in registry form, an `https` endpoint — is step 4's, with the rule's
  own reason.
- `wai-rs --features receipt_claims` (`crate::claim_receipt`, `wai_claim_vectors`) and
  `claim-conformance/`: vectors covering every reason, run by CI in both corpus loops,
  with the feature added to the library tests. A canonical fixture computed by
  a separate implementation (`claim_verify.py`: Python JSON with UTF-16 member order,
  `cryptography` Ed25519 plus step 3's key, `R` and `S` checks written from the curve's
  definition, BLAKE3 from its specification) is pinned in `wai-rs`; that script also
  reproduces every vector's outcome, reading OER receipts with the OER reference
  verifier. `claim-conformance/differential.py` runs both verifiers on generated inputs
  the corpus does not hold and fails on any differing verdict. Both are run by hand,
  not in CI.
- OER: `py_float_repr` wrote a double half-way between two shortest decimals with the
  upper one (`75.036163330078125` as `75.03616333007813`), where the reference's `repr`
  writes the even one (`…812`). §4's number profile keeps floats out of every receipt
  body, so no content address or verdict depended on it. `str()` of a number no check
  reads did — a vector's name, which the reference and `wai-rs` (`wai_oer_vectors`,
  `wai_oer_diff`) print — so the two printed different names for such a vector. The
  function now takes the length from the shortest round-trip formatter and the digits
  from the exact one, keeping the shortest digits where the nearest decimal of that
  length does not read back (next to a power of two); a table CPython generated pins
  fourteen such doubles.
- Not done here: key discovery (every claim signature presumes the verifier holds the
  key), an authority class for who may sign which class (a lawful-basis decision is
  checked against its declaration whoever signed it: nothing supplies an operator's key
  the way a verified delivery receipt supplies a publisher's), region- or time-scoped
  soft bindings, and checking a log's receipt under its profile.

### `wai.neural.int_mlicpp` — the MLIC++ entropy topology at decode-equivalence

- New capability `wai.neural.int_mlicpp` (`wai-rs/src/int_mlicpp.rs`, behind
  `neural_int`): the channel-autoregressive × checkerboard entropy model of
  `wai.neural.mlicpp` — `K` channel slices, each decoded as an anchor pass then a
  non-anchor pass, an entropy network over concat[latent decoded so far, hyper]
  re-run before every pass — with the hyper-decoder, the entropy network and the
  synthesis all integer networks (`int_hyper_synth`), the scale → distribution step
  `int_entropy::det_build_index`, and the integer rANS. Registry: `NeuralIntegerExact`
  (decode-equivalence), `Lossy`, royalty-free (its decoder is this crate's; the
  patent status of the techniques it implements was not examined).
- Why: the float capability's validation run decoded to garbage while its encoder
  evaluated the entropy network at its runtime's default thread count and the
  decoder single-threaded — one scale fell on the other side of a table edge at
  symbol 416 of 12,288, and every later symbol decoded wrong. Matching the execution
  settings fixes that run; it remains a convention. Here every scale is an integer
  both sides compute exactly, so no thread count, runtime or platform can move a
  bucket.
- Wire: a `WIM1` payload (36-byte header, then the hyper-latent and latent rANS
  streams) and a one-file `WMB1` model bundle (three `WHM1` networks and the entropy
  tables), so a pinned prior's SHA-256 covers everything the reconstruction depends
  on. The neural dispatch decodes it (`decode_envelope`, `decode_envelope_strict`)
  and checks a pinned prior against the very bytes it decodes with.
- Totality: every value the stream controls is range-checked. Before the decode
  allocates anything, each network's shapes are derived from its layer geometry (the
  new `int_hyper_synth::net_shape`, which shares its geometry with the checked run):
  a header whose shapes the model does not produce, or on which any network would
  form a buffer — an activation, a transposed convolution's uncropped plane — over
  `Header::MAX_TENSOR` elements, is refused as a model that does not fit the payload.
  `tests/int_mlicpp_preflight.rs` counts allocations: refusing such a header
  allocates a few kilobytes at most, where the decode without the check allocated
  the tensors the header asked for, and on a synthesis wider than the cap aborted
  the process on a 3.4-petabyte allocation. Latents go through
  the new `int_entropy::dequantize` (`sym · 2^frac + base`, `None` where `i64` cannot
  hold it; Kani harness `kani_dequantize_is_exact_or_refused`), and every network
  through the new `int_hyper_synth::run_net_checked`, which runs `run_net` layer by
  layer only while no product, partial sum or requantised value can leave its type
  and refuses otherwise. A test moves one channel's scales across table edges: the
  stream leaves the encoder at the first moved bucket, and the decode is refused
  when the desynchronised latents would overflow the entropy network. With the
  accumulation check removed, a damaged copy of the committed stream panics the
  overflow-checked build.
- Conformance vector `wai-rs/tests/vectors/int_mlicpp/` (from
  `tools/wai_int_mlicpp_export.py`, deterministic in its seed): **synthetic
  weights**, quantised like a trained model's; no trained MLIC++ model is used, so it
  shows decode-equivalence and nothing about compression quality. A 64×64 image, 8
  latent channels in 4 slices, all 16 scale buckets used, 47 of 512 residuals and 5
  of 16 hyper-latent symbols coded by the rANS escape. The export computes its
  reference in numpy integer arithmetic under explicit int64 range checks and decodes
  its own rANS bytes back, pass by pass, before writing. Pinned: the RGB (FNV-1a and
  SHA-256) and FNV-1a over the bucket and symbol sequences.
- Evidence: decoded to the pinned bytes on macOS arm64 and on wasm32-wasip2 under
  wasmtime; x86_64 not measured locally (determinism-tiers §6). A golden in
  `byte-exact-conformance` puts it on the four-platform matrix;
  `tests/int_mlicpp.rs` runs on every `wai-conformance` leg in both profiles, and the
  dispatch test in `wai-neural-gates`.
- Registry tests now pin a recorded basis for every `NeuralIntegerExact` capability
  (`NEURAL_INTEGER_EXACT_EVIDENCE`), as `STANDARD_DEFINED_EVIDENCE` does for
  `StandardDefined`.
- `int_hyper_synth::IntModel::from_bytes`, the compact network form the integer
  neural decoders load, is total on damaged bytes. It reserved room for the layer
  count it read before checking that the bytes held that many layers — about 480 GB
  for a count of 2^32 − 1, which aborts wherever that allocation fails — and formed
  layer sizes and read offsets in unchecked arithmetic, which overflows a 32-bit
  `usize` (on wasm32 the first case panicked with a capacity overflow). Both are
  checked now, and a valid model parses to the same layers.
- `int_hyper_synth::run_net` now applies each layer through one function that
  `run_net_checked` shares, `run_net_checked` takes each layer's geometry from the
  function `net_shape` uses, and `decode`'s pixel quantisation is `to_rgb8`; all
  refactors, and every committed vector reproduces its hash.

### Energy acquisition — the receipts wai-rs and wai-quantum seal can carry their class; meter windows open from rest

- [`extensions/energy-measurement.md`] §2.8: the world-session, film and reel,
  haptic, score, 4D-splat, audio-scene, ternary-reconstruction and
  provenance-step receipts, and every `wai-quantum` receipt, carry the
  acquisition class as an **optional signed label**. Absent, a receipt's signed
  bytes, identity hash and JSON are exactly the legacy ones, and the figure is
  unlabelled: accepted as legacy, never read as any class. Present, the signing
  domain's version byte becomes `0x02` and the class's §2.7 encoding follows
  the figure it labels, inside the signature. A label is refused as legacy
  `HwShunt`, beside a zero figure, or as a refined class without its evidence.
  Provenance labels each step, and a quantum bill of materials each entry.
- **Wire:** unlabelled receipts are unchanged; tests pin one receipt of every
  family to the bytes it had before the label existed, and every committed
  conformance corpus still verifies. JSON evidence fields are now
  `energy_`-prefixed (`energy_counter_filtering`, `energy_calibration_ref`) in
  every snake_case carrier, including the rendered `wai.video.int_motion`
  receipt: a quantum job receipt already renders `calibration_ref` for the
  calibration receipt it ran under. §2.7 now fixes every field name.
- §2.9: IOReport is classed `OnChipCounter` — the silicon's own reading, whose
  undisclosed derivation is part of the systematic error an `OnChipCounter`
  declaration already excludes. The section lists the documents elsewhere in
  the repository that class IOReport as `ModelBased` or `Estimator`, which this
  revision does not change.
- §2.3 and §3: the completion-count error depends on how the burst window
  opens. The reference meters read the count the moment their workers started,
  with no barrier, so the count read low by about `N/2` and every per-unit
  figure read **high** by about `N/(2C)` — the symmetric model did not describe
  it. Workers are now held at a start barrier and released as the window opens:
  the work lies in `[C, C + N)`, the figure divides by `C + N/2`, and the
  declared term is `N / (√3 · (2C + N))`. The window-opening mode is part of
  what a published figure reports.
- §2.3: a declaration describes the figure as sealed, and `joules_micro` is a
  whole number of microjoules, so the declaration includes the figure's own
  resolution: `0.5/√3` µJ for an estimate rounded to the nearest microjoule,
  `1/√6` µJ for the difference of two whole-microjoule register readings. The
  measurement's own terms are carried across in microjoules and the sum is
  divided by the figure sealed. A 0.563 µJ estimate measured to 1.54 % is
  sealed as `1` and declared 28.9 %. Every verifier refuses a hardware class
  with `joules_micro × relative_ppm < 288 676`, a declaration smaller than the
  integer's own rounding. The receipts built on `energy_class`, every
  optional-label family, the `wai-quantum` labels and the C2PA energy binding
  (bound or parsed) apply that check.
- §2.8: `quantum_energy::MaybeLabelled` parses and verifies a quantum receipt
  in whichever form it arrives. It takes the form from the JSON and refuses a
  malformed label rather than reading the receipt as unlabelled. The
  `wai-quantum verify` CLI and the HTTP handler's `POST /verify` use it, so the
  labelled receipt `wai_quantum_meter` emits for a measured figure verifies on
  both, and both restate the class. `MaybeLabelled::verify_against_calibration`
  checks a job's link to a calibration in either form, by the identity of the
  form the calibration is in; the `wai-web` job demo accepts a labelled
  calibration the same way.
- §2.2: the `OCMF` example now names what the regime rests on (a
  conformity-assessed, legally verified measuring system, not a calibration
  certificate); `OnChipCounter` is defined to cover any instrument whose
  uncertainty is not traceable, and the class is stated not to record where
  the sensor sits. §2.1 names the GREEN draft and its revision. §2 places the
  three-name mapping rule on producers. [`extensions/energy-binding.md`] and
  [`extensions/moq-streaming-format.md`] say "refined hardware class" where
  they meant it, so legacy `HwShunt` is not read as missing its evidence.
- Reference implementation: `energy_class` gains the label rules
  (`labelled_domain`, `write_label`, `check_label`, the JSON helpers) and the
  window-opening model (`CompletionCount`, `WindowOpening`). Each `wai-rs`
  family gains an `energy_provenance` field and `with_energy_class()`, which
  re-signs with the receipt's own key; `ProvStep` gains the same.
  `wai-quantum` adds `quantum_energy` without changing any existing type:
  `Labelled<R>` and `LabelledQbom` re-sign a sealed receipt over the labelled
  payload, and the inner receipt's own `verify()` refuses the result, so a
  reader unaware of labels fails closed. The new `energy_burst` module runs §3
  once for every meter; `wai_meter`, `wai_world_meter`, `wai_quantum_meter`,
  `wai_prov` and `wai_video_meter` now seal `OnChipCounter` with a declared
  uncertainty only when §4 and §2.3 allow, and seal unmetered otherwise. The
  world and quantum meters no longer count in batches, and the reported
  standard deviations are the sample ones the declared uncertainty uses. Each
  meter prints the resolution term and, when it seals unmetered, why.
- **Rust API (`wai`):** the receipt structs of the families above gain a
  public field, and `per_unit_relative_uncertainty` takes a `CompletionCount`.
  `wai_meter`, `wai_world_meter` and `wai_prov` now emit an unmetered receipt
  instead of exiting when no power reader is present.
- CI: a new `wai-energy-receipts.yml` runs the video/energy features, the
  label and pin tests of every family, `wai-quantum`'s own tests and the HTTP
  handler's verify test, on x86_64 and arm64. It builds four of the five
  reference meters (`wai_video_meter`, `wai_meter`, `wai_world_meter`,
  `wai_quantum_meter`). It does not build `wai_prov`, which needs the `codecs`
  feature (system libjxl); no workflow enables that feature. `wai-conformance.yml` now
  also triggers on `wai-quantum` changes, which its quantum corpora depend on.
- §2 and §2.8 now say what is enforced. The verifier rules are: a label is
  checked wherever one is present, and a receipt defined with a class from the
  start is refused when it carries a non-zero figure without one. Labelling a
  new figure in an optional-label family is a **SHOULD**, where it was a
  **MUST** that nothing checked and the reference implementation did not
  meet. An unlabelled figure sealed today has the bytes of an issued one, so
  no verifier can tell them apart. The reference implementation does not hold
  its producers to the rule either: in these families no constructor takes
  the class, and a label is added afterwards, by re-signing, or for a
  provenance step before its receipt is built. §2.8 says how a library could
  hold its callers to it: a labelled constructor, with the unlabelled path
  kept only for recreating issued receipts and deprecated, as `HwShunt` is.
  The reference implementation has not done that.
  The reference meters label every non-zero figure they seal. §2.8 names the
  two producers that do not: `world_model::frame_receipt`, which is not told
  its caller's classes, and the `wai-web` demos, which seal illustrative
  figures. It also says why a corpus generator that recreates a committed
  corpus, such as `wai_quantum_ops_vectors gen`, is not counted: every
  committed corpus predates the label.
- [`extensions/energy-measurement.md`] §1 states its scope: every receipt
  profile WAI defines that carries an energy figure, a list that is not
  closed. An OER/2 receipt is outside it. Its members are fixed, it has no
  member for the class, and its evidential axis is the grade. No rule of the
  extension binds an OER receipt or its producers and verifiers, so an OER
  receipt with a non-zero figure and no class breaks none, and the §4
  publication gate and §6 reporting contract do not reach it either. Only §2's
  rule for carrying a WAI figure into a three-name vocabulary reaches past the
  governed receipts. §1 also records that two receipts WAI defines fall short
  of §2: the JWP group receipt, which signs a non-zero `joules_total`, and the
  JWP object receipt, which carries `joules_micro`. Neither has a member for
  the class. Where a `WAI_RECEIPT` property accompanies an object over MoQ,
  the object's class travels in it, but no signature covers that property, so
  it falls short too. Nothing carries a class for a group total, and §2.8's
  rule that a total carries no class of its own does not cover it, because a
  group's parts are labelled outside the group receipt, if at all. Giving
  these a signed place for the class remains open.
  [`extensions/jwp-receipts.md`] §2.4.1 points to this from `joules_total`.
  [`extensions/interactive-worlds.md`] §8 now names the session receipt's own
  signing domain, `wai:world-receipt\x01`, which the reference engine already
  used.
- §2.6 says that the `wai.video.int_motion` and `wai.video.*` OTT constructors
  accept `HwShunt`, so that issued receipts can be recreated, and that nothing
  in them refuses it. The variant's deprecation warning reaches only code that
  names it; a class parsed from its label can be `HwShunt` without one. No
  producer in the reference implementation emits it.
- The resolution floor is stated as enforced: a declaration must reach the
  `0.5/√3` µJ (≈ 0.289 µJ) that rounding the figure to whole microjoules adds,
  `joules_micro × relative_ppm ≥ 288 676`. `SPEC.md`, the `wai-quantum`
  README, both crates' `BelowResolution` messages and several doc comments
  called it a 1 µJ floor, which is stricter than the rule verified.
  [`extensions/energy-binding.md`] §4 now states the floor, which `bind()`,
  `from_json()` and C2PA `recover()` already applied. The `energy_class` and
  `quantum_energy` docs say what the floor is: the rounding term every
  whole-microjoule figure carries. A figure read as a register delta carries
  the larger `1/√6` µJ, which its producer must declare and no verifier
  checks.
- [`extensions/oer-receipts.md`] §2 maps the evidential grade onto the refined
  classes: `CalibratedInstrument` and `OnChipCounter` support `measured`, and
  so does legacy `HwShunt`, for a figure sealed before the refinement or a
  refined figure mapped into the three-name vocabulary. Neither an OER/2
  receipt's `provenance.device` nor its `energy.method` names a class. The OER
  wire format, vectors and verifiers are unchanged.
- Not yet covered: the in-browser demos of `wai-web` that seal into an
  optional-label family seal illustrative figures, fixed in code or passed in
  by the page, without a label, and the OTT manifest demo labels a
  page-supplied figure `ModelBased`. The honest class of an illustrative
  constant is `Estimator`. The `ModelBased` and
  `Estimator` classes still carry no declared uncertainty.

### Energy acquisition — on-chip counters and calibrated instruments are separate classes

- [`extensions/energy-measurement.md`] §2 refines the hardware class. `HwShunt`
  covered both a counter the measured silicon keeps about itself (RAPL, NVML,
  SMC, IOReport) and an external meter, although only the second can carry an
  uncertainty traceable to a calibration. It is now two classes:
  `OnChipCounter` and `CalibratedInstrument`. Each **must** carry a declared
  uncertainty: a standard (k = 1) relative uncertainty in ppm plus the
  integration window it holds over. The counter records whether its output
  was filtered (`On` / `Off` / `Unknown`; RAPL energy filtering per
  INTEL-SA-00389), and the instrument records a reference to its calibration
  certificate. For an on-chip counter the declaration covers only what the
  implementation could evaluate, never the counter's own systematic error, so
  it is not a total-error statement.
- §2.1 keeps the declared uncertainty and the GREEN `data-source-accuracy`
  graded bands apart. A band bounds one device reading, while a declared
  uncertainty is a standard uncertainty about a per-object marginal figure. A
  band is never declared as a figure's uncertainty, and a declared uncertainty
  is never used to select a band.
- §2.3 and §3: a per-unit figure from the bracketed protocol declares the
  bracket's statistical term combined with a completion-count term,
  `N / (√3 · C)` for `N` workers and `C` counted completions. Up to `N`
  reconstructions are part-done at each edge of the window, and the count
  cannot see them. The thread count is now part of what a published figure
  must report.
- **Wire:** new signing tags `4` (`OnChipCounter`) and `5`
  (`CalibratedInstrument`), each followed by its evidence (§2.7). Tags `0`–`3`
  are unchanged and followed by nothing, so every receipt sealed with `HwShunt`,
  `ModelBased` or `Estimator` keeps its exact signed bytes and verifies as
  before. Tests pin those bytes. `HwShunt` is accepted when read, means
  "hardware, sub-class unspecified, no declared uncertainty", ranks below both
  refinements, and must not be emitted. JSON carriers (the C2PA companion
  assertion, the rendered int_motion receipt) add evidence fields only for the
  refined classes, so legacy bodies are byte-identical.
- Reference implementation: the new `energy_class` module is the one definition
  shared by `energy_binding`, `c2pa_emit`, `video_receipt` and the
  `wai.video.*` receipts. `verify()` and C2PA `recover()` reject a hardware
  class that lacks its uncertainty. `CounterMeter` emits `OnChipCounter` with
  the window it measured, declaring its configured relative uncertainty
  combined with the register delta's resolution term, `1/√6` µJ, so a short span declares
  more than the configured figure. `wai_video_meter` declares the bracket's statistical
  (Type A) term combined with the completion-count term, and emits an
  unmetered receipt when no decode completed in the window, where it used to
  count one.
- The receipt families that sealed `joules_micro` before the class existed now
  carry it as an optional label; see the next section.

### Integer decoders — fast kernels that keep every output integer

- `wai-rs` adds `int_kernels`: register-blocked convolution, transposed
  convolution and IGDN kernels (NEON widening multiply-accumulate on aarch64,
  portable Rust elsewhere), scoped worker threads (none on wasm), and a run-time
  `Policy` that selects them or the retained scalar reference kernels. The
  int_synth, int_hyper, int_audio, int_splat, int_world and JPEG-AI synthesis
  decoders dispatch to them. No bitstream, conformance vector or hash changes.
- The claim, precisely: for every input the reference decodes, every policy
  returns the same integer for every output element; with overflow checks on, a
  fast policy panics on the inputs where the reference panics, though the panic
  message can differ. Geometry the fast path does not handle, including a
  transposed convolution whose sizes the reference's own arithmetic cannot form,
  goes to the reference. The fast path's own allocations and loops stay in
  proportion to what the reference reads and writes, whatever the stride or
  padding. The two paths still allocate different buffers, so an allocation that
  fails is outside the claim. It is argued by construction and tested, not
  machine-proved. The rules, the tests, their limits and the measured throughput
  (one machine) are in
  [`research/decode-throughput.md`](research/decode-throughput.md).
- CI: `wai-conformance` runs the kernel-equivalence tests, every committed
  vector that reaches the kernels and edge geometry through the public decoders
  (strides up to `usize::MAX`, padding far beyond the kernel, and strides and
  paddings at which the reference's size arithmetic overflows), under every
  policy, with overflow checks on and off, on each platform leg. Geometry on
  either side of the padded-input bound is pinned too: at the bound the fast
  kernel runs, beyond it the reference does, and a test-build count of
  fast-kernel entries shows which one ran.
  `wai-byte-equality` and `wai-formal-verification` fire on changes to
  `int_kernels.rs`.

### Headline decode tests fail on a missing input instead of passing

- The `int_hyper` tests over the pretrained mbt2018_mean (`int_hyper_synth`,
  `int_video_sota_keyframe`) and the JPEG-AI whole-decoder tests
  (`jpegai_e2e_full`, `jpegai_reconstruction_binding`, `jpegai_live`) read inputs
  from outside the tree, and used to `return` — reported as a pass — when those
  were absent. With
  `WAI_REQUIRE_SOTA_VECTORS=1` a missing input is now a failure that names the
  tool producing it; without it the skip is written to the process's real stderr,
  so it shows in an ordinary `cargo test` run. Either way, a corrupt input file
  fails, and so does a variable that names inputs which are not all there: a
  path that does not exist (`WAI_JPEGAI_CODESTREAM`/`_WEIGHTS_DIR`/`_NORM`, even
  when another of the three is unset), or a directory that does not exist or
  lacks a file the test reads (`WAI_SOTA_INT_DIR`, `WAI_JPEGAI_ASSETS`, the
  JPEG-AI weights directory). So without the switch a test skips only when no
  variable names what is missing: one of the three path variables is unset, or
  `WAI_SOTA_INT_DIR` or `WAI_JPEGAI_ASSETS` is unset or empty and its default
  directory (`/tmp/sota_int`, `../jpegai-vm-assets`) is absent or incomplete.
  The CI jobs always set the switch, so the new test
  `require_vectors_rule` checks this rule without it: it runs the helpers in
  child processes with the switch and the variables set per case, and
  `wai-sota-vectors.yml` runs it.
- New CI job `wai-sota-vectors.yml` regenerates the mbt2018_mean integer vectors
  from pinned inputs (`tools/wai_sota_vectors.sh`, the Python stack in
  `tools/requirements-sota-int.txt`, a public-domain photograph pinned by SHA-256,
  the published checkpoint pinned by its full SHA-256) and runs those tests
  with the switch set. It is triggered by changes to the decode path and by a
  weekly schedule. Besides requiring each test by name, it fails if the test log
  contains a skip banner, since a skipped test also reports `ok`. The export tool
  checks the checkpoint's bytes before loading them (as tensors only), because the
  model zoo's own download path checks no hash. The job checks the Rust decode
  against the Python integer reference on one runner; cross-architecture
  byte-equality remains the committed tiny vector's job in
  `wai-byte-equality.yml`. The regenerated vectors are not portable, and part
  of them is not repeatable run to run on every machine. Between an arm64
  macOS machine [name removed 2026-09-30] and the x86_64 runner, the synthesis-only test's input latent,
  which the float model computes, differed, and with it its expected
  `rgb_hash`; each side wrote the same latent in all of its runs. Two runs on
  the one runner wrote different `cond.json` and `int_synth_sota.json`, most
  likely in the float decode's PSNR, which only the quality test reads, against
  thresholds. The PNG bytes of the source also differed between the Mac and the
  runner; the PNG encoder writes them from the pinned pixels before any model
  runs, and the pixels matched. So no expected hash of the regenerated vectors
  is pinned in the repository (the committed tiny vector's is, in
  `tests/vectors/int_hyper/tiny.json`); each run checks the hashes its own export
  wrote, and a hash from one run's vectors must not be pinned for another run,
  on the same machine or another. The pins that hold across machines are on the
  inputs: the photograph's pixels and the checkpoint's bytes.
- New test `integer_decode_keeps_the_float_codec_quality`: the Rust decoder
  measures its own reconstruction against the source pixels and must stay within
  1 dB of the same model's float decode on the same image. On the pinned
  photograph, measured on an arm64 macOS machine [name removed 2026-09-30]: integer decode 35.985 dB (8-bit
  PSNR of the Rust decode), float decode 36.175 dB (the export tool's PSNR of the
  float reconstruction), a cost of 0.19 dB, from 6328 bytes of rANS payload for
  256×256 pixels. This is a relative bar on one image: nothing here compares
  mbt2018_mean with another codec, and quality is image-dependent — the 42.6 dB
  figure quoted for this codec was measured on a different photograph, which is
  not in the repository.
- New CI job `wai-jpegai-e2e.yml` runs `jpegai_e2e_full` and
  `jpegai_reconstruction_binding` with the switch set, with the same by-name and
  skip-banner checks. Their inputs — the decoder
  weights, normative tables and test codestream, extracted from the JPEG-AI
  reference-software assets (which are not in this repository) — are fetched from
  WAI's versioned models root (`jpegai/v1/`) by
  `tools/wai_jpegai_fetch_e2e_inputs.sh` and checked against the SHA-256 manifest
  `tools/jpegai-e2e-v1.sha256`. The job also pins the SHA-256 of the decoded
  pixels (`WAI_JPEGAI_EXPECT_SHA256`): measured natively on an arm64 macOS machine [name removed 2026-09-30]
  and reproduced by an x86_64 Linux build under QEMU user-mode emulation, and
  required again on the x86_64 runner. That extends a measurement, not a
  registry guarantee: the registry classes `wai.image.jpegai` as
  `NeuralEntropyExact`, a conservative class that guarantees only the entropy
  decode to be byte-identical on every machine, while the exact decoder's pinned
  output for these inputs was measured identical on four platforms
  (determinism-tiers §6).
- `jpegai_live` (the ONNX-runtime `wai.image.jpegai` decode) follows the same
  missing-input rule, but no CI job runs it: its model bundle exists only in the
  local JPEG-AI asset directory, and it needs the `neural` and `codecs` features.
- determinism-tiers §5–§6 record the two jobs. The full-scale `int_hyper` row and
  the JPEG AI whole-decode row (split from the index-derivation row, whose test CI
  still does not run) now name their workflow, state that a missing input fails
  there instead of skipping, and state that neither workflow has run through the
  forge: each passed under `forgejo-runner exec` on the Linux runner's host in an
  earlier form, and its current steps have run on macOS arm64 only. §5 quotes the
  `int_hyper` cost measured on the pinned photograph in place of the 42.6 dB
  figure, and names the bar each job re-checks.

### Prior carriage and derived-track operations — registered with their limits

- New [`extensions/prior-carriage.md`]: a `Prior` MAY declare the wire form of
  the artifact its digest covers (`prior.form`, additive on the wire; manifests
  without it round-trip byte for byte). When declared, the digest covers the coded
  bitstream, and a sink that cannot read the form refuses the prior rather than
  substitute another packaging. `wai.prior.nnc` (ISO/IEC 15938-17:2024) is the
  registered form. It is optional, not recommended: royalty status is not
  established, and whether conforming decoders reconstruct identical tensors is
  not established, so a `replicate` cannot rest on a prior read out of a form not
  listed as exact (none is). As a payload, NNC takes the replicate class of a
  standard whose decoder conformance the registry has not checked, as MPEG-I
  Haptics does; its own conformance clauses were not read. `wai.tensor.nnc` is
  registered as a candidate for general tensors; its coding is in an unpublished
  amendment, which
  is tracked, not pinned. The extension also fixes the precedence of in-band
  data-unit signatures: WAI's digest covers every byte, and an in-band verdict is
  recorded beside it, never instead of it.
- New [`extensions/derived-tracks.md`]: the ten derivation operations of
  ISO/IEC 23001-16:2021 are registered as `wai.derive.*` strings, each with a
  per-operation determinism tier. Dissolve and Scaling are untiered. No operation
  is exact unconditionally, Identity included. An executed instance is exact only
  when no sample value is converted into another representation or computed, its
  output does not change over the span the claim covers, and its operation's own
  conditions hold; the sink lists each condition it checked. A derived sample
  holds the tier of its weakest step. A claim over it names the operations
  actually executed, the parameter values each applied, the span it covers and
  its inputs, visual and parameter. These are not payload capabilities, and WAI
  does not composite. Clause 8 of the standard, which defines each operation, was
  not read. Every tier rests on conditions checked on the executed instance; the
  geometric conditions listed per operation come from the operation's title, as
  does whether it reads a visual input, and reading clause 8 may show a list to be
  incomplete. Clause 4 crops or fills an input of another size implicitly for
  every operation, and no list carries a condition on where such an input is
  placed.
- `wai-rs`: `codecs::derivation` (registry, `Condition`, `derived_sample_tier`),
  `codecs::DeterminismTier` (the two tiers as values, re-exported by
  `codecs::derivation`), `codecs::prior_form`, `Prior::{new, with_form, form,
  check_form, form_preserves_exactness}`, and `container::replicate_prior_refusal`,
  which states why a `replicate` is refused, including over a declared form.
- **Behaviour change:** `Prior::is_cross_hardware_exact` now also requires the
  declared form, if any, to be established exact (`form_preserves_exactness`), so
  a prior that declares a form is not cross-hardware exact whatever its contract
  says; `determinism_class` still reads the contract alone. A prior without a form,
  which is every prior written before `form` existed, gets the same answer as
  before. The rule sits in that predicate because every `replicate` check reads
  it, `container::replicate_is_sound` and `Wai::replicate_is_sound` included.
- Each `ExecutedOp` names the visual inputs its instance read (`reads`, as an
  entry of `inputs` or the output of an earlier executed operation), because
  clause 4 of ISO/IEC 23001-16 lets an operation that is not first read the
  output of any previous operation, only new inputs, or both, so its position
  does not say what it read. An instance of an operation that reads a visual
  input and names nothing gives no tier, wherever it sits: it read an input that
  nothing names, whether a track, item, track group or preceding output the file
  lists or, where nothing lists one, the default fill picture (clause 5.1). When
  at least one operation executed, a call that contradicts itself (a reference to
  an unlisted input or to an operation that does not precede the reader, or a
  listed input no instance names) is a `DerivedSampleError`; an empty executed
  list gives no tier before those checks. `neural::ModelRegistry::resolve_checked`
  refuses a declared form the sink cannot read, and the reference sink reads
  none. A new CI workflow, `wai-neural-gates`, is configured to run the two
  model-free tests of those neural-sink refusals, which no workflow ran before,
  in an Ubuntu 24.04 container image; it has not yet run on a runner.
- **Rust API, source-breaking.** `container::Prior` gained a public field,
  `form`. Its fields are all public and it is not `#[non_exhaustive]`, so Rust
  code outside the crate that builds a `Prior` with a struct literal, or
  destructures one without `..`, stops compiling. The fix is to add
  `form: None`, or to build it with the new `Prior::new(id, sha256,
  determinism)` (and `.with_form(form)`), which names no optional field and so
  is unaffected by the next optional field added. Reading fields, serde, and
  functional update (`Prior { form, ..other }`) are unaffected. SPEC §8 versions
  the wire format, not this API. On the wire, `form` is an optional key, added
  as `prior` itself was, and absent it the manifest's bytes are unchanged. SPEC §8
  also says the manifest schema is frozen for all of 1.x; how an optional key is
  added under that freeze is not settled here, for `form` or for `prior`. For the
  crate, Cargo's semver rules treat a source-breaking change in a 0.x crate as
  needing a new minor version: the release that carries this must be 0.4.0 or
  later (`wai-rs/Cargo.toml` still says 0.3.0). `#[non_exhaustive]` was not
  added. It would break the same literals now and, in addition, rule out the
  `form: None` fix and functional update from outside the crate, which the
  crate's own integration tests use. It would also make `Prior` the only
  manifest struct that code outside the crate cannot build with a literal; the
  others are `Manifest`, `ModelRequirement`, `Conditioning`, `ManifestV2` and
  `RenditionMeta`. Whether they should all become `#[non_exhaustive]`, so that
  adding an optional key never breaks Rust callers, is a decision for all of them
  together.
- Open: the operations' four-character codes are not recorded (clause 8 of
  ISO/IEC 23001-16 was not read); the `wai` CLI does not list derivation
  operations; there is no NNC reader.

### Broadcast and streaming codecs — registered, licence-classified, never defaults

- SPEC §5 registers the elementary video and audio bitstreams distribution chains
  carry: `wai.video.{av2,vvc,hevc,avc,vp9,evc,evc.baseline,avs3,lcevc}` and
  `wai.audio.{mpegh,avs3,aac}`. Each names one payload framing, which §8 then
  fixes forever: the Annex B byte stream for AVC, HEVC, VVC and LCEVC (each
  standard also defines a bare NAL unit stream, which is not the capability),
  AV2's Annex B length-delimited bitstream, IVF for VP9, the Annex B raw
  bitstream file storage format for EVC, MHAS for MPEG-H 3D Audio and ADTS for
  AAC. The AVS3 video and audio rows name each standard's elementary
  bitstream; those texts were not read, so whether either defines a second
  form is unchecked. Each also carries a licence class with its recorded
  basis; two are classifications rather than verified statuses (AVS3: no
  published fee schedule; EVC Baseline: royalty-free by design, no formal
  guarantee). All are sink-supplied; none is in the floor or the recommended
  set. No decoder and no C dependency is added.
- AV2 is registered royalty-free but not recommended: the reference decoder
  exists, deployable decode is not yet confirmed. Promotion is a MINOR revision.
- LCEVC is an enhancement layer, carried embedded in its base bitstream or as a
  `WAI2` companion whose `companion_of` must name one of its base codecs (AVC,
  HEVC, VVC, EVC, VP9, AV1, under every capability naming those bitstreams;
  `codecs::enhancement_bases`) and resolve to exactly one selectable rendition
  of it; a base takes at most one enhancement layer. The base list is WAI's
  registration, not a restriction of ISO/IEC 23094-2, which leaves the base
  decoder unspecified; a later revision may add a base (a widening, §8).
- `wai.video.lcevc` and the HDR metadata descriptors (`wai.meta.hdr10plus`,
  `wai.meta.dovi`) are companion-only, now stated normatively: never a v1.0
  `capability` or `fallback`, never a v1.1 selectable rendition. `pick` /
  `primary` never return them; `Wai::check_component_roles` (v1.0) and
  `WaiMulti::check_component_roles` (v1.1) report envelopes that break the
  rules (`RoleError`, `#[non_exhaustive]`). A `WAI1` naming one only as its
  `fallback` does not conform and never takes dispatch step 5, but still decodes
  via its primary capability (§4 step 3); new `Wai::dispatch_capability` applies
  §4 with that rule, and `wai decode` and `wai inspect` both use it, so they
  agree on such a file. `wai wrap` refuses to write a non-conforming envelope
  and, with `--companion <cap> <payload>`, writes the `WAI2` companion form,
  whose intent follows the same rule as the `WAI1` form's (`--intent`, below),
  checked over the base payload's capability and each companion's. HDR
  metadata may bind to a neural video payload, and the base check is what
  refuses `wai wrap wai.neural.dcvc_rt … --companion wai.meta.hdr10plus …`
  (`NeuralFloat`, SPEC §7) unless another intent is given. Every companion-only
  capability is in a standard class, so the companion check refuses nothing
  yet. `wai unwrap --component <i>` extracts a `WAI2` component.
- SPEC §7.1 states the three rules a `WAI2` companion meets, and
  `WaiMulti::check_component_roles` enforces each: (1) its capability is
  companion-only or unregistered — a registered standalone capability marked
  companion is `RoleError::StandaloneAsCompanion`; (2) `companion_of` names a
  capability it applies to (`codecs::companion_applies_to`: for HDR dynamic
  metadata a registered capability in the `video` media class — new
  `codecs::is_video_class`: every `wai.video.*` capability that is not
  companion-only, the neural video codecs `wai.neural.{video_bmshj2018,dcvc_rt,
  dcvc_fm}` and the films `wai.film.{linear,reel}` — or an unregistered one,
  whose class is unknown; for LCEVC a base codec) —
  `RoleError::IncompatibleBase`; (3) that base is present, and unique for an
  enhancement layer. The check previously let an HDR descriptor bind to an audio
  rendition, LCEVC bind to any `wai.video.*` rendition (JPEG 2000 included), and
  any standalone codec ride as a companion; the reference test that carried an
  IAMF stream as a companion now carries ST 2094-10 dynamic metadata [name removed 2026-09-30] instead.
- SPEC §8 now says what these rules are, so they do not read as a redefinition
  inside `1.x`: a registration fixes a capability's payload format **and its
  role** (standalone media, or companion-only); changing either, or narrowing
  what a companion-only capability applies to, needs a new MAJOR. A MINOR
  revision may add a rule that enforces a registration, or a manifest field's
  definition, as it was made, if it lists every kind of envelope it makes
  non-conforming. The companion rules are such rules: the HDR descriptors were
  registered as companion descriptors to a video payload, not standalone media;
  `wai.video.lcevc` was registered companion-only with these rules; and
  `role = "companion"` was defined as naming, in `companion_of`, the component
  it binds to. SPEC §7.1 lists the envelopes they make non-conforming: an HDR
  descriptor as a `WAI1` `capability` or `fallback`, an HDR descriptor as a
  `WAI2` selectable rendition (which `pick` / `primary` no longer select), an
  HDR descriptor bound to a registered capability outside the `video` media
  class, a registered standalone capability marked companion, and a companion
  of any capability with no `companion_of` or none that names a selectable
  rendition in the envelope. Such envelopes still parse. The list also records
  that the earlier §7.1 named immersive-audio metadata as a companion use and a
  reference test carried `wai.audio.iamf` that way, which rule 1 now rejects,
  since `wai.audio.iamf` was registered as standalone media.
- SPEC §8 also says what registering a capability does to an envelope that used
  its string while no revision registered it. §7.1 leaves such a use unchecked
  (an HDR descriptor's unregistered base is not constrained; an unregistered
  companion needs only a present base), so the registering revision may make the envelope
  non-conforming where it uses the string against the new registration: a
  companion whose capability is registered standalone, an HDR descriptor bound
  to a capability registered outside the `video` class, a companion-only
  capability carried as a payload or bound to a base it does not admit. §8
  lists these kinds once for every registration; they still parse. This covers
  `wai.video.lcevc`, which no earlier revision registered.
- `WaiMulti::companions_of` no longer says every companion it returns is meant
  to be applied: the role check accepts an unregistered companion with a
  present base, and SPEC §7.1 forbids a sink to apply a companion whose
  capability it does not recognise.
- The broadcast-codec dispatch guidance asks for a recommended-set `WAI2`
  rendition, not a bare v1.0 `fallback`: step 5 decodes the same payload with
  the fallback's codec, which cannot read an HEVC, VVC or AV2 bitstream.
- `codecs::capability_license` is now exhaustive with no default class, so a
  registered capability can no longer become royalty-free by omission. New
  `codecs::mandatory_floor()` / `recommended_set()`; tests hold that every
  default is royalty-free. `wai capabilities` gains a `role` column.
- **Replicate classes: exhaustive, and the exact class needs recorded evidence.**
  `codecs::capability_replicate_class` no longer has a default arm (the
  determinism-tiers change below still ended it in a `StandardDefined`
  fall-through). Every registered capability is listed under one class; one
  listed nowhere returns `None`, which the tests reject. `StandardDefined`, which
  promises the same bytes on every conforming sink, is admitted only on a recorded
  basis — a conformance clause that requires identical output, or the WAI
  extension whose equivalence pins the reconstruction — pinned one row per
  capability in the tests (`standard_defined_rests_on_recorded_evidence`). The
  broadcast codecs enter with:
  - `StandardDefined`: AVC, HEVC and VVC (ITU-T H.264 §C.4, H.265 and H.266
    §C.5.1: every sample a conforming decoder outputs equals the specified
    decoding process's) and VP9 (VP9 specification v0.6 §8.1);
  - `StandardTolerance`: AV2 (AV2 v1.0.0 §7.1: film grain, as AV1);
  - `StandardUnclassified`: EVC (both entries), AVS3 video and audio, LCEVC,
    MPEG-H 3D Audio and AAC, whose conformance documents were not read.

  `wai.audio.ivas` moves from `StandardUnclassified` to `StandardTolerance` on
  3GPP TS 26.252 V19.2.0 §7.2.1: implementations based on the floating-point
  reference are tested by non-bit-exact procedures. New
  `codecs::is_cross_decoder_exact(cap)` is true where the class's
  `replicate_promise` is `ByteExact`, except for a capability whose output is
  computed from other envelopes' decoded output (new
  `codecs::composes_decoded_sources`: `wai.audio.scene`). A scene's mixdown is
  exact given its sources' PCM, and a source decoded through `wai.audio.opus` may
  decode differently on two conforming sinks, so the capability alone does not
  make it portable; new `codecs::is_cross_decoder_exact_with_sources(cap,
  sources)` is true once every source's capability is known and exact, and
  false when it cannot tell (no source given, or a source outside
  `wai.audio.*`). The class's `ByteExact` promise for a scene holds given its
  sources' PCM, so SPEC §7 names `is_cross_decoder_exact`, not the promise, as
  the audit for the same bytes on every sink. The
  promise covers the output the standard or extension specifies, not a
  conversion a sink applies afterwards. New
  `tools/wai_replicate_probe.sh` decodes each format with every independent
  decoder on the host and fails if a `StandardDefined` capability's decoders
  disagree — a falsification check, not the basis of any class.

### JPEG AI exact decoder: no platform math library on the path, and a cross-platform check

- The exact decoder (`jpegai_decode::decode`) called the platform math library's
  `exp` in the per-channel gain scaler, on every target. Its roundings to an integer
  compiled, on x86_64 and wasm32 (as rustc 1.98 compiles them), to calls to the
  symbols `round`, `roundf` and `rint`, which the link resolves; an x86_64 Linux
  test executable inspected took them from Rust's compiler builtins, not from the C
  library. The decoder now takes `exp` and rounding from the new `float_repro`
  module, which builds them from IEEE-754 basic operations, integer conversions and
  bit manipulation, so their bits are fixed by the source. `float_repro::exp` was
  within 1 ULP of the host math library on every host it was compared with
  (determinism-tiers §6 lists them). Its test asserts ≤ 2 ULPs. The test is ignored
  on targets whose C library was not measured, Windows among them. It selects by
  target, not by host, so it also runs on glibc versions and on x86_64 hardware that
  were not measured. The roundings equal the standard library's. The decoded output
  is unchanged (RGB8 SHA-256 `dbc494a5…` for the `jpegai/v1` inputs).
- The exact decoder now also refuses a target that evaluates float arithmetic in x87
  extended precision (32-bit x86 without SSE2), where it would not reproduce.
- New `jpegai_decode::decode_with_stages`: the same decode, also reporting the float
  values after each stage (latents, synthesis, each post-filter), so two platforms
  can be compared stage by stage rather than only by their final bytes.
- New `tests/jpegai_cross_platform.rs` pins SHA-256 hashes of the parsed float
  inputs, of every reported stage and of the RGB8 output for the `jpegai/v1` inputs.
  All nine matched on macOS arm64 (native), wasm32-wasip2 (a WASI runtime), Linux
  arm64 (glibc) and Linux x86_64 (glibc, under user-mode emulation), each running
  this change's final source, and on all four also before the math-library calls
  were replaced. determinism-tiers §6 records the toolchain and the method for each.
  Not measured: macOS x86_64, Windows, x86_64 hardware. The test skips without its
  inputs; `wai-jpegai-e2e.yml` supplies them and requires it to pass; `float_repro`'s own pinned-output test
  needs no inputs and runs in the library tests.
- Two checks keep math-library calls off the decode path. Clippy's
  `disallowed_methods` (`wai-rs/clippy.toml`), denied in each module on the path,
  rejects a listed float method however it is spelled.
  `tests/jpegai_decode_symbols.rs` lists the symbols of an executable that links
  only the decoder and its weight loaders, and fails on any C math-library function
  the executable imports, whatever made the call, and on any it defines other than
  the round-to-integral functions, whose results IEEE 754 fixes (unix targets). A
  definition alone shows what the linker kept, not what the decode calls: GNU ld
  kept an unused `floor` in an x86_64 Linux debug build. The conformance workflow
  runs both on its unix legs. They replace a scan of the source text, which missed a
  call spelled `f64::exp(x)`. Details and evidence: determinism-tiers §5–§6. The
  registry class of `wai.image.jpegai` is unchanged here.
- The gain unit's `log_k` constant was `0.2003654931615394`, 9.2 × 10⁻⁹ from
  `(ln 54.82 − ln 0.11) / 31`, which the reference computes as
  `0.20036548400207613`. It is now that value. At β = 0, the only β the exact decoder
  accepts, both constants give the same scaler for every gain-table entry, so the
  decoded output is unchanged.
- `tools/wai_jpegai_gain_exp_margin.py` now models the reference's gain scaler as the
  reference computes it, in float32, rounding half to even. It fails if the decoder's
  `log_k` differs from the reference's. At β = 0 every table entry is at least 77
  float32 ULPs from a rounding boundary, so the two scalers agree for any float32
  `exp` within that. The margin does not hold over every 12-bit β: at four inputs
  the reference's own scaler depends on the last bit of its `exp`
  (determinism-tiers §6).

### OER receipts — graded, chained energy receipts with three implementations

- New [`extensions/oer-receipts.md`]: ingests OER/2 draft 0.1. A receipt records one
  unit of work, its joule cost, the cost's **evidential grade** (`projected`,
  `simulated`, `derived`, `measured` — orthogonal to the acquisition class, mapping
  stated), and what it rests on; canonical, content-addressed, hash-chained and
  Ed25519-signed. Conformance is reason-set-equivalence plus a canonical-form fixture
  the vectors cannot check themselves.
- New `oer-conformance/`: the upstream 9 vectors, 1 from the portable C kernel
  (`settle/`, Apache-2.0), 99 hand-declared edge vectors — every reason the upstream
  nine never exercise, each number shape and decimal-string defect and each limit
  from the accepting side, each bound, each of the four tolerances at and past its
  edge on both sides and past 34 digits, each step of the order of the reasons
  reported alone, each member judged on its value, each claim an abstention cannot
  make, each key a signature cannot rest on, and the extremes a document may hold —
  and `unreadable/`, the documents every verifier must refuse whole.
  They found five defects in the reference verifier (a mis-quoted
  `missing_capability`, three crashes on missing members, a repeated receipt id
  hiding reasons), and a class of inputs on which the reference and `wai-rs`
  reported different reason sets or crashed (parser-dependent numbers and versions,
  lenient decimals, members of the wrong type, rounding past the 28th digit,
  documents at the edge of what a parser can read); all fixed in both.
- `wai-rs --features oer` (`crate::oer`, `wai_oer_vectors`): a third, independent
  verifier; CI runs it on the four-platform matrix and rebuilds the C kernel,
  requiring its output byte-identical to the pinned corpus.
- `oer-conformance/fuzz_differential.py` with `wai_oer_diff`: seeded differential
  fuzzing of the reference verifier against `wai-rs` over mutated corpus documents;
  any divergence in reasons or refusal, or any crash, fails it.
- Departures from draft 0.1, each in the extension's §9: decisions are named
  `oer.decision.*`, not `wai.settle.*` (WAI's namespace is for payload formats); the
  stage list behind `bundle.provenance_root` is serialised exactly as the body is,
  pinned by a non-ASCII fixture; the number profile is a MUST (I-JSON integers,
  decimal strings in JSON's number grammar, anything else `malformed`); every
  composition check is exact, with `ln 2` pinned as a decimal, and every tolerance
  scales by the figure's magnitude; a document outside the document profile is
  refused whole, a repeated member name included; every mistyped member has one
  named reason; published keys are decoded strictly, so a small-order key cannot
  verify every message; `oer` is a string;
  and a decision that was not earned is receipted as `abstain` — no answer, no
  record — which a new reason, `abstention_settled`, holds to that. `escalate` is a
  verdict returned from upstream, as the corpus always had it.

### Determinism tiers — naming the reproducibility spectrum for neural decode

- New [`extensions/determinism-tiers.md`]: names the two guarantees a neural
  capability can make, so a manifest can declare which it satisfies and a verifier
  knows what it may check. `entropy-consistency` — the entropy decode reproduces
  across devices (integer entropy arithmetic or explicit parameter transmission);
  JPEG AI sets its bit-exact conformance point at this tier (the entropy-decoder
  output). `decode-equivalence` — the whole decode to the final medium reproduces
  on every conforming sink, so the rendered result has a portable hash; WAI's
  integer capabilities are registered at this tier, and it is the precondition for
  a hard reconstruction binding. Referenced from SPEC §7 (Replicate-soundness);
  framed self-referentially.
- Normative: over a neural capability, `intent = replicate` promises
  `decode-equivalence` (SPEC §3, §7; standard codecs, below). The
  first version of the extension called `entropy-consistency` the floor for a
  sound neural `replicate`, and SPEC §7 said the same in its closing paragraph
  while requiring an integer decode path in its rule. Resolved: a neural
  capability backs `replicate` only if it is registered at `decode-equivalence`,
  which §7 admits only by the integer construction; `entropy-consistency` MUST NOT
  back it, and neither may a reconstruction that one implementation pins through
  fixed-order float stages. A tier is a property of the capability as registered,
  not of one implementation.
- **Downgrade: `wai.image.jpegai` is registered at `entropy-consistency`, not
  `decode-equivalence`, and no longer backs `replicate`.** Its registry class
  moves from `StandardDefined` to the new `NeuralEntropyExact` (integer entropy
  decode, float later stages). Its decode has float stages and a float ONNX
  fallback, so decode-equivalence never held as a property of the capability, and
  `StandardDefined` ("the neural float hazard does not apply") did not describe it.
  What the exact decoder does is unchanged — one pinned reconstruction, reproducible
  across machines under the conditions in the extension's §5 — and is now stated as
  a property of that decoder. A reconstruction binding over `wai.image.jpegai` is
  checkable only with that decoder (reconstruction-binding §2). Earlier entries
  below that call the JPEG AI decode decode-equivalent, or byte-identical on every
  machine, are superseded by this one.
- **`replicate` over a standard codec promises the named standard's decoder
  conformance, not byte identity across sinks.** The first form of this change
  stated in SPEC §3 that `replicate` promises the same bytes on every conforming
  sink, while every standard codec stayed `StandardDefined` and passed the rule —
  including codecs whose standard conforms decoders to a bound around a reference
  (RFC 6716 §6 holds an Opus decoder to `opus_compare` thresholds, not to
  identical output). SPEC §3 and §7 now state the promise per registry class, and
  two classes are added:
  - `StandardTolerance` — the named standard holds decoders to a criterion around
    a reference: `wai.audio.opus`, `wai.image.jpeg`, `wai.image.jxl`,
    `wai.image.jp2`, `wai.video.j2k`, `wai.video.av1`, `wai.video.av1.lossless`,
    `wai.image.avif` (AV1 §7.2: identical output unless film grain is signalled,
    and then no "perceptually significant differences", a criterion the AV1
    specification leaves undefined), `wai.audio.iamf`.
  - `StandardUnclassified` — the named standard's conformance clauses have not
    been checked: `wai.audio.ac4`, `wai.audio.ivas`, `wai.mesh.draco`,
    `wai.pc.gpcc`, `wai.pc.vpcc`, `wai.splat.sog`, `wai.haptic.mihs`,
    `wai.meta.hdr10plus`, `wai.meta.dovi`, and the pass-through asset formats
    `wai.mesh.glb` and `wai.splat.spz`.

  `StandardDefined` keeps the capabilities whose reconstruction is specified
  exactly (PNG, FLAC, zstd, XZ and the WAI-native capabilities). `replicate` over
  the reclassified capabilities is still honoured; what it promises is narrower.
  SPEC §7 also states that a conversion a sink applies after the standard's decode
  (PNG depth reduction, YUV→RGB, JPEG colour space, rendering an IAMF mix) is not
  part of the promise, and that an extension composing other capabilities' output
  is only as exact as their classes (spatial-audio §4 now says so for sources
  decoded through Opus). The clauses read for each classification, and what was not
  read, are in `research/media-codec-registry.md`. API: new
  `ReplicateClass::{StandardTolerance, StandardUnclassified}` variants,
  `ReplicatePromise` and `ReplicateClass::replicate_promise`; `wai capabilities`
  prints `standard-tolerance` / `standard-unclassified` for them.
- `wai.image.jpegai` stays `NeuralEntropyExact` as a conservative choice pending a
  reading of its conformance part, ISO/IEC 6048-4. An earlier form of this change
  said in SPEC §7 and the extension that its standard does not bound the
  reconstruction, so no weaker promise than `decode-equivalence` was available.
  Part 4 had not been read, and the overview of the standard that was read (arXiv
  2510.13867) says the conformance part "specifies the requirements for generating
  a JPEG AI-compliant reconstruction" and "considers the possibility that a decoder
  is conformant to the standard without having bit-exact reconstruction", with "a
  certain leeway" — which points to a criterion around a reference, though the
  overview does not state one. SPEC §7 (*`wai.image.jpegai` and the standard
  classes*) now states what was and was not read, that the capability is a
  candidate for `StandardTolerance` if Part 4 sets such a criterion, and that its
  treatment differs from the `StandardUnclassified` codecs' and film-grain AV1's
  by this revision's choice, not by anything found in the standards. No class or
  runtime behaviour changes.
- `decode_envelope_strict` and `Wai::replicate_is_sound` now apply SPEC §7 in two
  steps (`codecs::replicate_refusal`): the capability's registry class first, then
  the pinned prior. Before, they read only the prior's *declared* contract string,
  so a `replicate` envelope over `wai.image.jpegai` or a float codec passed when its
  prior declared an integer contract; it is now refused. API: new
  `ReplicateClass::NeuralEntropyExact` variant (an exhaustive `match` must add it),
  `DeterminismTier`, `ReplicateClass::{neural_tier, label}`,
  `codecs::{replicate_refusal, ReplicateRefusal}` and `Wai::replicate_refusal`;
  `Wai::replicate_is_sound` now consults the registry (a floor capability with no
  prior now passes it; the free function `container::replicate_is_sound` is
  unchanged and checks only the prior).
- Fix: `wai wrap` wrote `intent = replicate` over every capability, including
  `wai.image.jpegai` and the float neural codecs, which SPEC §7 says an encoder MUST
  NOT do, and it pins no prior, so its envelopes over an integer-exact neural codec
  were refused by the strict decode as well. It now takes `--intent <intent>`
  (default `replicate`) and writes `replicate` only where a sink accepts it with no
  prior — over the three standard classes; otherwise it refuses, says why, and
  writes nothing. The same holds for a `--fallback` capability, since a sink that
  lacks the primary decodes through it (SPEC §4 step 5): `wai wrap wai.image.jp2 …
  --fallback wai.image.jpegai` is refused unless given another intent. SPEC §7 now
  says the class a sink checks is that of the capability it decodes through. API:
  `codecs::unpinned_envelope_intent`. The library test checks every registered
  capability against the sink's own check; `tests/cli_wrap_intent.rs` runs the
  binary, which needs the `codecs` feature, which no CI job builds.
- CI: new `wai-neural-gates.yml` runs the two model-free neural-sink tests,
  `strict_decode_gates_unsound_replicate` (the `decode_envelope_strict` gate) and
  `resolve_checked_verifies_prior_hash`, on the Linux runner, and fails unless both
  ran. No workflow ran them before: they need the `neural` feature, which
  `wai-conformance.yml` does not build, and `wai-neural-e2e.yml` selects only the
  reference-match tests. The job runs in an Ubuntu 24.04 image: the ONNX Runtime
  binary the `ort` crate downloads needs glibc 2.38 and a GCC 13 libstdc++, and
  in the runner's default 22.04 image the test binary fails to link.
- Wording corrections, each checked against its source:
  - The claim that byte identity holds between sinks running the same decoder
    build (the `StandardTolerance` and `ReplicatePromise` docs, SPEC §7) had
    nothing behind it and is removed; `replicate` over those classes does not
    promise byte identity across sinks.
  - JPEG and JPEG 2000 were listed with Opus and JPEG XL as having a stated bound.
    The clauses that would state them were not read (T.81 leaves the IDCT's
    accuracy to its Part 2; only the table of contents and terms of ISO/IEC
    15444-4 were read), so SPEC §7, the extension and the registry docs now say so
    and give no bound.
  - The JPEG AI overview is quoted instead of paraphrased: it says Part 4
    "considers the possibility" of conformance without bit-exact reconstruction,
    and does not itself state a criterion on the reconstruction.
  - "Before v0.4.0" in the registry comments and "until v0.4.0" in the extension
    were wrong: the replicate classes and `wai.image.jpegai` were both added within
    v0.4.0. They now say "earlier in v0.4.0".
  - determinism-tiers §2 described ISO/IEC 23001-11's decoding-operation reduction
    as in-band metadata with which a decoder degrades its output after a conformant
    decode. The 4th-edition CD text (WG 3 N1864) says otherwise: a decoding
    operation reduction request goes from the receiving device to the remote
    encoder, which sends a less complex bitstream, and complexity metrics let a
    decoder vary its operating frequency; neither changes the output of a decode.
    §2 now sorts every metadata set that text's Introduction names by what it
    changes. Attenuation-map post-processing, display adaptation (the RGB
    components of reconstructed frames rescaled, saturated or, for contrast
    enhancement, set to zero or the peak signal before display) and cross-segment
    decoding (an enhanced picture used as a reference) change the decoded medium,
    and the post-decode rule applies to them. Decoding-operation-reduction and
    attenuated-video requests, and low-power encoding, change the payload before
    it arrives; the rule about "honouring a hint" is replaced by one about such a
    payload: a tier, hash or receipt over it covers that payload, not the one it
    stands in for. Media selection and complexity metrics change neither. §2 also
    says what the metadata is: power-saving metadata stated in advance, whose
    energy quantities are expected or potential savings (clauses 7.4 and 8), not
    energy that was spent. energy-binding §7.1–§7.2 and task-fidelity §4 follow;
    §7.2 now names the parties to `op_requested` and `op_committed` as the
    requesting and the acting party, where "sender" and "receiver" had them
    reversed for a DOR request. The transport landscape research note marks its
    earlier reading as superseded in a note under its provenance header that
    names every passage repeating it.
  - `int_entropy.rs` cited SPEC §6 for replicate-soundness; it is §7.
- Correction to the JPEG AI entries (SPEC §5 row and the decoder entry below): the
  dequantisation, latent refinement, post-filters (including the eICCI network) and
  colour conversion run in IEEE-754 float with a fixed evaluation order, not in
  integer; and the end-to-end test bar is max|Δ| ≤ 2 (recorded max|Δ| = 1), not
  1 LSB.
- Fix: `wai.image.jpegai` decoded codestreams it does not implement. The exact
  decoder checked only the picture size and latent channel counts, so a 256×256
  codestream with β ≠ 0, another `model_id`, another synthesis transform or another
  eICCI model selection passed that check and was decoded as if it had the
  implemented configuration; the float ONNX path was no better (it applied β to the
  entropy parameters but not to dequantisation). The earlier statement that every
  per-image quantity is derived from the bitstream did not hold while β is not
  implemented; β ≠ 0 is now refused. `jpegai_decode::check_codestream` now
  checks every header field either path hard-codes, and both paths refuse a
  codestream that fails it. The exact decoder also refuses to run in a non-default
  floating-point environment (`fp_env_is_default`). The committed reference
  codestream (`wai-rs/tests/vectors/jpegai/`) lets the library tests exercise the
  check without fixtures.
- Fix: the JPEG AI picture-header parser mislabelled a flag. The bit after each
  component's GRFS channel flags is `synthesis_tile_enable`; the LSBS flags are the
  first two bits of the tool header (`parse_lsbs_flags`). It also read `c_ver` /
  `c_hor` when `s_ver` / `s_hor` is 2, where the reference codes neither; the float
  path's parser shared that defect and dropped the separate chroma β.
- Float reproducibility stated precisely (SPEC §7, the extension, `DeterminismClass`):
  IEEE-754 basic operations are correctly rounded, and float decodes diverge across
  hardware through what a runtime chooses per device and through the floating-point
  environment. `replicate` still requires integer arithmetic, because those
  conditions belong to one build and one process.
- The extension now also states what a receipt can hash under each tier, why
  transmitting entropy parameters stops at `entropy-consistency`, which capabilities
  are offered at `decode-equivalence` and at what cost, and the evidence behind each
  claim — which tests skip, which CI does not run, and which matrix legs actually
  ran (the Windows x86_64 leg has not run since 2026-08-27).

### Reproducible-reconstruction binding — a C2PA/JPEG-Trust hard binding over pixels

- New `wai.*` provenance capability: `reconstruction_binding` (feature
  `provenance`) + [`extensions/reconstruction-binding.md`]. A cryptographic hard
  binding over the *deterministically decoded* pixels of a neural-coded asset —
  which such media normally can't carry, because a neural decode isn't
  reproducible (JPEG AI's normative conformance point is the entropy-decoder
  output, not pixels; its multi-branch design reconstructs one codestream several
  ways). WAI's byte-exact decode restores a portable decoded-pixel hash. SHA-256
  (the C2PA default) over the canonical RGB8 buffer plus a descriptor pinning the
  reproducible decode; carried as a custom C2PA/JPEG-Trust assertion
  (`science.transaction.wai.hash.reconstruction`) beside the standard byte-binding.
  A substrate for machine-readable AI-content marking (e.g. EU AI Act Art. 50).

### `wai.image.jpegai` now dispatches to the byte-exact decoder

- `decode_envelope` prefers the deterministic `jpegai_decode::decode` when the
  model directory ships the integer weight bundle (`weights/*.json` + `norm.json`),
  falling back to the ONNX-runtime decode otherwise — existing bundles keep
  working. `wai.image.jpegai` is now registered in SPEC §5 (it was a live
  capability missing from the normative menu).
- The cross-arch byte-equality matrix now covers the merged decoder's entropy
  stage (`byte-exact-conformance`), and that crate's build — broken on a moved
  `rans.rs` include — is restored.

### JPEG AI — a byte-exact codestream→RGB decoder (`jpegai_decode::decode`)

- `wai::jpegai_decode::decode(codestream, &DecoderWeights, &NormTables) -> Rgb`
  reconstructs an ISO/IEC 6048 JPEG-AI codestream to RGB through a fully
  integer entropy front end and a deterministic value/synthesis/post-filter
  path — one pinned reconstruction, reproducible byte-for-byte on every machine,
  so the decoded pixels have a portable hash. No per-image fixtures: every
  per-image quantity is derived from the bitstream; only checkpoint constants
  come from a sink-supplied model bundle.
- Determinism profile: the entropy decode (mANS), the dequant, and the derived
  entropy inputs (skip mask, σ index, RVS band, dequant gain) are integer-exact
  against the reference; the PTQ value/synthesis stages are deterministic and
  land within 1 LSB of the ISO reference decoder. Distinct from the in-house
  `wai.neural.int_*` codecs (all-i64, hash-portable): this decodes a published
  external standard's codestream deterministically.
- Modules: `jpegai_decode` (the composed entry point) over `jpegai_bitstream`,
  `jpegai_mans`, `jpegai_pipeline`, `jpegai_synthesis`, `jpegai_icci`
  (feature `neural_int`). Inputs are regenerable via
  `tools/wai_jpegai_build_e2e_inputs.py` from the JPEG-AI reference-software
  assets, which are kept locally and are not in this repository; the extraction is
  published under WAI's versioned models root (`jpegai/v1/`). The whole chain —
  codestream → RGB — is exercised end to end against the reference PNG by
  `wai-jpegai-e2e.yml`, which fetches those inputs pinned by SHA-256.

### JPEG AI — a live, self-sufficient `wai.image.jpegai` decode (2026-06-22)

- `neural::jpegai::decode(codestream, &JpegAiModel)` decodes a real ISO/IEC 6048
  codestream to RGB with **no per-image schedule fixtures**: every per-image
  quantity is derived from the bitstream, and only checkpoint constants come
  from a sink-supplied model bundle. Wired into `decode_envelope` — JPEG AI is
  now a dispatchable capability, not just reference/conformance support.
- The two derivations that closed it, traced bit-exact from the reference: the
  residual coding mask is `sigma_skip > 382` (the σ surface *before* the rvs
  addition — on the bucketed index it looks ambiguous, on `sigma_skip` it is
  exact, 0 violations / 65536 positions); the dequant is
  `resi × rvs_dq[grfs[ch]][lk] / 2^16 / scaler_vec[ch]`, with `grfs` determined
  the same way the σ rvs does so the per-channel `scaler_vec` is consistent.
  Matches the reference decoder's PNG to within rounding (max diff 1), checked by
  `tests/jpegai_live.rs` on a machine holding the local JPEG-AI assets; no CI job
  runs it (its model bundle is not in the repository).

### Film reel — `wai.film.reel`, cuts over one world (2026-06-17)

- The editor's layer over `wai.film.linear`: a `WREL` embeds one sealed world
  + a sequence of **shots** (each its own camera track + frame schedule) played
  back to back. Cutting wide→close, or one beat from two angles, is film
  grammar; a reel carries it as a few kilobytes of camera keys, the world
  shared across every shot. Conformance is **reel-equivalence** —
  film-equivalence extended over cuts.
- Shots MAY revisit earlier world ticks (a cut back in time from a new angle),
  so the world is sampled at the sorted/de-duplicated union of all frame ticks
  — convergence still holds because replay-equivalence is order-independent.
- Reuses the film machinery (shared camera interpolation + pose bytes) + the
  `FilmReceipt` profile (identity = `reel_hash`); `wai_meter` grows a `.wrel`
  branch. Registered + tiered Lossless. Corpus vector `coverage` (the character
  beat cut three ways, incl. a cut back in time) joins `film-conformance/`,
  verified by `wai_film_vectors` on the four-platform matrix. SPEC §5 row +
  `extensions/linear-film.md §7`.

### Linear film — `wai.film.linear` (2026-06-16)

- "A film is a sealed, camera-on-rails world." A generated film ships not
  as pixels but as the **world** that produces it (an embedded sealed
  `wai.world.replay`) + a **camera track** (keyframed pose over the world's
  ticks) + a **frame schedule** (which tick each frame samples). The sink
  renders the frames — lever 3 (instructions-at-the-sink) for the
  synthesizable slice of "video" (animation / CGI / motion graphics). The
  conformance corpus carries whole films in ~1 KB.
- `wai-rs` `film` feature (reuses the world engine + `Fx` floor + blake3):
  the `WFLM` container (contract + embedded `WWLD` + camera track) and the
  deterministic reconstruct contract. Conformance is **film-equivalence** —
  per-frame `BLAKE3(world_state_hash(frame_tick) || interpolated camera
  pose)` is bit-identical on every machine. It **composes** the embedded
  world's replay-equivalence (inherited unchanged) with the camera track in
  the `wai.det.fixed64` floor; the render (pixels) stays the sink's.
- The world engine gained one general primitive: `collect_checkpoints_
  sampled` / `run_sampled_dispatch` — read the state hash at an arbitrary
  frame cadence, independent of the contract's checkpoint cadence (a
  renderer samples the timeline at its own frame rate).
- Corpus `film-conformance/` + `wai_film_vectors gen|verify` (dolly,
  two-shot, crane-up) — all pass, on the four-platform CI matrix. Registered
  + tiered Lossless (film-equivalence is exact). `FilmReceipt` (per-frame
  Merkle leaves) + a `wai_meter` `.wflm` branch emitting **measured** joules.
- Spec PROMOTED to `extensions/linear-film.md`, SPEC §5 temporal-content
  row, content-taxonomy updated (animation-film → ✅). The biggest
  commercial wedge in the taxonomy, now shipped.

### Music as instructions — `wai.audio.score`, LIVE at /score (2026-06-14)

- The deepest audio lever: a score carries **no recording**, only note
  events — ship the composition, not the recording. `wai-rs` `score`
  feature (reuses the world `Fx` floor + blake3): the `WSCR` container
  (contract + note events) and a **trig-free** deterministic reference
  synth — a phase accumulator drives a piecewise-linear oscillator
  (saw/square/triangle/pulse) shaped by a linear ADSR, all `+ − × ÷`.
  Pitch is a per-sample phase increment authored off-line (no exponential
  in the floor). `mixdown_hash` is the **mixdown-equivalence** quantity,
  the music counterpart of spatial audio; a soundfont/neural instrument
  is the sink's presentation.
- Corpus `score-conformance/` + `wai_score_vectors gen|verify` (scale,
  chord/polyphony, timbre-changing arpeggio) — all pass. Registered +
  tiered Lossless (exact reference mix).
- Browser sink `ScoreSink` (`wai-web`) + `/score` page, LIVE: synthesizes
  the score in-tab, verifies the mixdown hash vs the corpus, draws a
  piano roll, and plays the reconstructed PCM via Web Audio. wasm pkg
  uniquely named `wai_score`.
- Spec PROMOTED to `extensions/score.md` (site mirror), SPEC §5 row,
  content-taxonomy updated (music → ✅). Completes the audio surface:
  recorded objects (audio.scene) and synthesized notes (audio.score).

### Neural subsystem reconciled onto main (2026-06-14)

- Wires the full learned-image/audio decoder family into `neural/mod.rs`
  and brings the categorical `rANS` (`decode_categorical`) it needs:
  hyperprior (`mbt2018_mean`), checkerboard context (`elic`),
  channel-autoregressive (`mlicpp`/`mlicv2`/`cmic`), integer-exact
  (`int_synth`), `glc` (VQ), inter-frame video (`dcvc_rt`/`dcvc_fm`),
  `snac` audio — plus the ISO/IEC 6048-1 JPEG-AI entropy + codestream
  + σ/value/filter port (`jpegai_*`).
- Restores the **semantic / generative path** (`semantic.rs`,
  CLIP-similarity conformance, `ModelRequirement.min_sim`/`sim_metric`)
  and the shared deterministic ONNX loader (`ort_session.rs`), behind
  the `semantic` feature.
- Re-adds `byte-exact-conformance` (cross-arch rANS + int_synth + JPEG-AI
  mANS goldens) and the model export/encode tooling under `tools/`.
- All features build + test green together (`neural` + `semantic` +
  the interactive/temporal cargo classes); clippy clean.

### Asset provenance — `wai.asset.provenance`, LIVE at /provenance (2026-06-13)

- **Lever 4 — the "verifies whatever it carries" half.** For content that
  cannot be reduced to instructions (a photo, a recording), WAI does not
  reconstruct — it carries the opaque bytes content-addressed and attaches
  a signed, energy-accounted **lineage**. `AssetReceipt` (`provenance.rs`,
  `provenance` feature): a JWP group whose objects are processing
  **steps** (op + input/output content hash + work + measured joules),
  Merkle-bound, Ed25519-signed, with chain-linkage verification (each
  step's input = the previous output, final output = the asset hash) and
  `parent_receipt_hash` chaining. `from_json` so a verifying sink can
  re-parse it.
- **`wai_prov` tool**: synthesizes a real image, builds a 3-step lineage
  (origin → PNG master → published JPEG) with the **real measured CPU
  energy** of each encode (IOReport via macmon, no sudo), wraps the JPEG
  as a content-addressed `wai.image.jpeg` envelope, and emits the signed
  receipt (origin 0 + PNG 0.43 J + JPEG 0.14 J = 0.57 J measured).
- **Browser sink — `AssetProvenance` (`wai-web`)** + `/provenance` demo,
  LIVE: recomputes the asset's content hash in-tab and compares to the
  receipt (tamper-evidence — a "corrupt one byte" button flips it red),
  verifies the Ed25519 signature + Merkle chain, shows the energy-
  accounted lineage, and hands the carried JPEG to a native `<img>` (the
  browser decodes it; WAI does not). wasm pkg uniquely named `wai_prov`.
- **Spec PROMOTED to `extensions/provenance.md`** (site mirror). It is a
  receipt profile *about* a carried asset (like `jwp-receipts`), not a §5
  payload codec — the asset keeps its own capability (`wai.image.jpeg`).
  Completes the thesis: carries whatever it can reconstruct, verifies
  whatever it carries.

### Cargo-class polish wave — real cloud, EWA render, receipts, telepresence (2026-06-13)

Four follow-on builds across the new cargo classes:

- **Real-SPZ-derived 4D clip** (`wai-spz-gen 4d`): the `bloom` clip animates
  1200 genuine SPZ [name removed 2026-09-30] Gaussians (the same cloud the static
  `wai.splat.spz` asset ships, round-tripped through the SPZ library) —
  real cloud, authored motion. LIVE on `/volumetric`. Honest framing: the
  win is a verifiable per-frame state, not capture fidelity.
- **WebGPU EWA render on `/volumetric`**: the page now renders the
  reconstructed cloud with a real anisotropic EWA splat pipeline (the
  same one the worlds demo uses) — per-Gaussian covariance projection,
  conic, 3σ bounding quad, premultiplied blend, back-to-front sort — with
  a 2D alpha-disc fallback where WebGPU is unavailable. The hash-verify
  badge is render-independent. Confirmed on WebGPU (a browser [name removed 2026-09-30]) + 2D fallback.
- **Per-class receipts with measured joules**: `HapticReceipt`
  (`haptic/receipt.rs`) + `Splat4dReceipt` (`splat4d/receipt.rs`) — the
  worlds/audio JWP receipt profile transposed to touch (per-channel
  blocks) and volumetric (per-frame blocks), reusing `world::receipt::
  merkle_root`. New `wai_meter` bin measures **real CPU energy** (IOReport
  via macmon, no sudo) for haptic/splat4d reconstruction and emits a
  signed, self-verified receipt carrying the measured `joules_micro`;
  `wai_world_meter` now also dispatches `feed0`/`skeletal0`, so feed and
  avatar worlds get the same measured `SessionReceipt`. Every cargo class
  now has a measured-energy receipt path (haptic 57 µJ, splat4d 0.21 J for
  1200 Gaussians × 17 frames, feed/avatar mJ-scale).
- **Avatar telepresence — LIVE at `/telepresence`**: a live skeletal world
  of N avatars over the existing worlds relay. New `SkeletalActor`
  (wasm) signs `CREATE` + `SET_ROTATION` `CausalOp`s; the worlds transport
  (relay-first + BroadcastChannel fallback, hello/sync/op/hash,
  boundary-hash convergence) is reused verbatim. Each tab creates and
  animates its own arm rig; sinks converge on the pose hash ("poses agree
  @ tick N"). Verified on prod through the real relay with two isolated
  contexts — identical pose hashes, zero errors.

### Volumetric / 4D splats — `wai.splat.4d`, LIVE at /volumetric (2026-06-13)

- **Phase 5 complete — the roadmap frontier.** A `wai.splat.4d` is a
  **temporal Gaussian splat**: the static splat extended to time. Each
  Gaussian carries a keyframed track (position / scale / rotation /
  color / opacity); the sink interpolates the cloud at frame N in the
  `wai.det.fixed64` floor. Conformance is **frame-state-equivalence**:
  the reconstructed Gaussian set at frame N hashes identically on every
  sink, no tolerance — the temporal counterpart of the static splat.
- **`wai-rs` `splat4d` feature** (reuses the world `Fx` floor + blake3,
  no codec/ML deps): the `W4DS` container (contract + per-Gaussian
  keyframe tracks, 14 `Fx` per key) + a deterministic reconstruct (Fx
  linear interpolation, hold outside the track). Trig-free (quaternions
  interpolate component-wise; the sink normalizes for rendering).
  `frame_hash` = the frame-state-equivalence quantity; `clip_hash` =
  frame-independent identity. Registered + tiered **Lossy** (consistent
  with `wai.splat.spz` — the splat representation approximates the
  captured scene, even though the reconstruction is frame-state-exact).
- **`volumetric-conformance/` corpus** + `wai_splat4d_vectors gen|verify`:
  orbit (two Gaussians swap places), pulse (scale + fade), morph (color
  shift) — every frame's hash checked, all pass.
- **Browser sink — `Splat4dSink` (`wai-web`)** + `/volumetric` demo page,
  LIVE: reconstructs the cloud in-tab and **verifies every frame's hash**
  against the corpus; a frame scrubber re-derives the cloud, drawn as
  painter-sorted alpha Gaussians (a sink with WebGPU renders the same set
  with the full EWA pipeline — presentation). wasm pkg uniquely named
  `wai_splat4d`.
- **Spec PROMOTED to `extensions/volumetric.md`** (site mirror),
  registered in SPEC.md §5 (`wai.splat.4d`). **Honest about size**: the
  win is content-addressing + dedup + a verifiable per-frame state, NOT
  payload reduction (a real cloud is MB-scale) — stated plainly.
- **The cargo set is complete across the medium**: sight (worlds/replay),
  hearing (audio/mixdown), touch (haptics/sample), live state
  (feed/replay), a moving body (avatar/pose), and now a moving volume
  (splat.4d/frame-state) — one `wai.det.fixed64` floor, one verification.

### Avatars / digital humans — `wai.avatar.pose`, LIVE at /avatar (2026-06-13)

- **Phase 4 complete — digital humans**: motion as instructions. An
  avatar is a **skeletal world** — the worlds container/op-log/ordering/
  sealing/state-hash/receipt-chain, with the `step` reduced to the
  identity. Conformance is **pose-equivalence** (the per-bone local
  transforms hash identically), a special case of replay-equivalence.
- **`wai.world.sim.skeletal0`** (in `wai-rs` `world` feature): a bone is
  a `parent[16] + translation[3] + rotation-quaternion[4] + scale[3]`
  (all `wai.det.fixed64`). Mutations: `CREATE` / `DELETE` /
  `SET_ROTATION` (the animation driver) / `SET_TRANSLATION` / `SET_SCALE`,
  all LWW. **Trig-free floor**: rotations are quaternions and the FK
  composition is `+ − ×` only; axis-angle authoring happens off-line.
  The forward-kinematics → world space and the skinning of a
  `wai.mesh.glb` are presentation (a pure function of the hashed pose),
  never in the hash. Registered + tiered `Lossless`; dispatched in
  `run_sealed_dispatch`.
- **`avatar-conformance/` corpus** + `wai_avatar_vectors gen|verify`: a
  sealed 3-bone arm wave, a sealed 5-bone humanoid pose, and three seeded
  adversarial-arrival shuffles of the arm wave — all converge, all pass.
- **Browser sink — `SkeletalSink` (`wai-web`)** + `/avatar` demo page,
  LIVE: replays a sealed animation in-tab, **re-derives the sealed pose
  hash** (checked against the native corpus and the file's own seal), and
  exposes FK world-space joints + bone segments to draw the skeleton on a
  rotating orthographic canvas with a tick scrubber. wasm pkg uniquely
  named `wai_avatar`.
- **Spec PROMOTED to `extensions/avatars.md`** (site mirror), registered
  in SPEC.md §5 (`wai.avatar.pose`). A moving body joins sight (worlds),
  sound (audio), touch (haptics), and live state (feeds) under one floor.

### State-feeds — `wai.feed.state`, LIVE at /feed (2026-06-13)

- **Phase 3 complete — live state-feeds**: state as instructions. A
  feed is a **non-physics world** — the same `WWLD` container, op log,
  canonical ordering, sealing, state hash, and receipt chain as
  interactive-worlds, with the simulation `step` reduced to the
  identity. Conformance is therefore **replay-equivalence**, reused
  wholesale: no new criterion.
- **`wai.world.sim.feed0`** (in `wai-rs` `world` feature): nodes are
  records of self-describing typed cells (`Int` i64, `Fixed` Q32.32,
  `Text` UTF-8 — hashed without a schema). Mutations: `CREATE` /
  `DELETE` / `SET_CELL` (LWW) and `ADD_INT` / `ADD_FIXED` (SUM —
  commutative accumulation). Wrong-typed / out-of-range / absent-node
  updates are deterministic no-ops. Registered + tiered `Lossless` in
  the central `codecs` registry; dispatched in `run_sealed_dispatch`.
- **`feed-conformance/` corpus** + `wai_feed_vectors gen|verify`:
  sealed scoreboard (SUM scoring + an LWW correction), telemetry
  (LWW gauge + SUM packet count, incl. a negative `ADD_FIXED`), market
  tape (LWW last + SUM volume), and three seeded adversarial-arrival
  shuffles of the scoreboard — all converge, all pass (hash equality).
- **Browser sink — `FeedSink` (`wai-web`)** + `/feed` demo page, LIVE:
  loads a sealed feed, replays it in-tab with the same engine that made
  the corpus, and **re-derives the publisher's sealed BLAKE3 state hash**
  (the file's seal also equals the native corpus hash — both checked).
  A tick scrubber re-derives the feed at any tick; the record table
  renders the typed cells generically. wasm pkg uniquely named
  `wai_feed` (avoids the deploy fingerprinter's basename collision).
- **Spec PROMOTED to `extensions/state-feeds.md`** (site mirror),
  registered in SPEC.md §5 (`wai.feed.state`). The **broadcast profile**
  (one authoritative publisher, many read-only sinks) is documented as a
  deployment shape of `wai.world.live`, not a new format. State joins
  sight (worlds/replay), hearing (audio/mixdown), and touch
  (haptics/sample) under one floor — verified the same way.

### Interactive worlds — `world` feature + conformance corpus (2026-06-11)

- **New cargo class implemented**: the interactive-worlds extension
  (spec draft staged in `wai-transaction-science-web/research/`,
  destined for `extensions/interactive-worlds.md`). Media that consumes
  input: a world is a content-addressed artifact, every event is a
  signed `CausalOp`, and conformance is **replay-equivalence** —
  identical BLAKE3 state hash at every checkpoint, no tolerance
  parameter. One container (`WWLD`) serves `wai.world.scene` /
  `wai.world.replay` / `wai.world.live`, differing only in op-log seal
  state; an input log and a multiplayer op log are the same object.
- **`wai-rs` `world` feature** (pure Rust, no ML dep chain; deps:
  blake3 + ed25519-dalek): `Fx` Q32.32 fixed-point (`wai.det.fixed64`,
  the floor profile — bit-exact unconditionally), xoshiro256++ contract
  rng, signed/dep-gated ops with canonical `(effect_tick, lamport,
  actor)` ordering, coordination-free BLAKE3 node minting, WWLD
  section container with content-binding seal, `Sim` trait (the
  extension standardizes the contract, not a physics engine) +
  `kinematic0` reference registrant, sealed `Replayer` + `LiveWorld`
  engines (buffering, past-tick re-derivation, equivocation rejection,
  receipted-checkpoint history sealing).
- **`world-conformance/` corpus** + `wai_world_vectors gen|verify`:
  sealed replays, LWW/SUM conflict batteries, dep-gap / past-tick /
  shuffled-arrival live streams, expected hashes. All vectors pass;
  clippy clean on the new code.
- **Spec PROMOTED to `extensions/interactive-worlds.md`** (site mirror
  `src/pages/extensions/`), registered in SPEC.md §5: `wai.world.scene`
  / `wai.world.replay` / `wai.world.live` + asset media `wai.mesh.glb`
  (canonical glTF 2.0 Binary, Khronos — decoder sink/library-supplied);
  `media: "world"` noted in §3. Capability strings live in the central
  `codecs` registry (`known_capabilities()`), tiered `Lossless` —
  replay-equivalence is exact state reproduction, the simulation analog
  of lossless.
- **Second sim registrant `wai.world.sim.ballistic0`** (gravity +
  ground-plane bounce, per-node radius/restitution, `G = −10`, all
  Q32.32 — bit-exact through every bounce) proving the registration
  story: entered registry, dispatch, and corpus (`ballistic_bounce`
  vector) without one line of the extension changing. New
  `run_sealed_dispatch` routes on `contract.sim`; a missing sim
  reports the world inert (`UnknownSimCapability`), mirroring SPEC.md
  §4 step-6 semantics so callers can fall back per the envelope.
- **Native energy meter — measured joules, no sudo (2026-06-13)**:
  corrected a wrong assumption that real energy needed RAPL/sudo
  (unavailable on the arm64 prod box). On arm64 macOS [name removed 2026-09-30] the SoC energy
  counters are readable through IOReport with **no privileges** (the
  source `macmon`/`asitop` use; verified euid≠0 reads CPU/GPU/ANE power
  on the arm64 machine [name removed 2026-09-30]), and `proc_pid_rusage` gives exact per-process
  retired cycles/instructions. New `wai_world_meter` bin (`world`
  feature): brackets a tight replay burst of a sealed `.wwld` against an
  idle baseline, samples CPU power via macmon, and reports the marginal
  **measured** CPU energy — e.g. the truss replay metered at ~512 µJ /
  replay (~4.3 µJ/tick) at 3.2 W marginal, a real number that fills a
  receipt's `joules_micro` where a native sink runs. The burst runs on
  N threads so the marginal draw clears a shared box's idle fluctuation
  (single-thread 512 µJ vs 10-thread 524 µJ/replay cross-validates the
  method — per-replay energy is invariant to parallelism). With a second
  arg the meter emits a full §8 `SessionReceipt` carrying the MEASURED
  joules + exact `work_steps`, signed and self-verified — real measured
  energy end to end. Spec §8 + the `/worlds` page corrected: the
  sink-metering path needs no sudo (IOReport on arm64 macOS [name removed 2026-09-30], RAPL on
  x86); the browser estimate is the fallback, not the ceiling. clippy
  clean, 94 tests green.
- **Phase 2 begun — haptics (`wai.haptic.signal`) (2026-06-13)**: touch
  as instructions. Spec drafted (`research/haptics-draft.md`), capability
  registered + tiered. `wai-rs` `haptic` feature (reuses the world `Fx`
  floor, no codec/ML deps): the `WHAP` container + a deterministic
  reconstruct — per channel, a keyframed **intensity/sharpness** envelope
  (the Core-Haptics pair every actuator API maps to) linear-interpolated
  in `Fx` and quantized to `u8`. Trig-free by construction (amplitude +
  sharpness, not a modulated carrier), so the floor reconstructs it
  exactly with `+,−,×,÷` alone. `signal_hash` is the **sample-equivalence**
  quantity. `wai_haptic_vectors gen|verify` + `haptic-conformance/` (ramp,
  thud, dual-rumble pulse) — all pass, hash equality. 104 tests, clippy
  clean. Next: Gamepad/Vibration browser sink + `/haptics` demo, then
  promote. This completes the sight → hearing → touch trio's engines.
- **Spatial-audio: receipt profile + spec PROMOTED (2026-06-13)** —
  Phase 1 complete. `audio_scene::SceneReceipt` (the worlds receipt
  transposed to audio): identity `scene_hash`, the JWP group's objects
  are the mix's **PCM blocks** (`BLAKE3("wai:audio-block")`), Merkle-bound
  with JWP's exact domains, one Ed25519 sig over scene_hash + root +
  blocks + `work_mac` (samples × objects, exact signed) + carried
  `joules_micro`; `parent_receipt_hash` chains renders. Reuses
  `world::receipt::merkle_root` — minimal new code. 2 receipt tests (101
  total). Draft promoted to `extensions/spatial-audio.md` (+ site mirror,
  SPEC §5 row + extension pointer) — and the canonical spec corrected to
  the **as-built** `sqrt_constpower` pan (sqrt-only, no transcendental
  tables) and the real receipt shape, since registration froze the
  format. clippy clean; spec page live. Phase 1 (spatial audio) DONE:
  spec · register · mixer · corpus · live /audio sink · receipt ·
  promoted. Next: Phase 2 — haptics.
- **Spatial-audio browser sink — LIVE at /audio (2026-06-13)**: `wai-web`
  `AudioSink` (wasm-bindgen): parse `WSCN`, load the corpus sources, run
  the exact deterministic mixdown, expose `mixdown_hash_hex` +
  `scene_hash_hex` + `mix` (i16 PCM) + `object_xz(t)`. The `/audio` page
  (site-styled, in nav) loads a scene, **verifies the mixdown hash
  in-browser against the corpus** (mixdown-equivalence, the same proof
  worlds does for state), plays the bit-exact stereo mix via an
  AudioBuffer, and shows a top-down radar of the objects panning. Proven
  on prod: all three scenes verify, zero errors — wasm in a browser [name removed 2026-09-30] mixes
  byte-identically to native. Honest split kept: plays the exact stereo
  floor; binaural/HRTF noted as the richer optional presentation.
- **Spatial-audio conformance corpus (2026-06-13)**: `wai_audio_vectors
  gen|verify` + `audio-conformance/` — the mixdown-equivalence analog of
  the worlds corpus. Three vectors (static stereo pair, distance
  near/far, keyframed left→right sweep) ship a `.wscn` scene + a
  `.sources.bin` (decoded mono tones) + expected `scene_hash` and
  `mixdown_hash`; `verify` mixes each and checks hash equality, no
  tolerance. All pass; clippy clean.
- **Spatial-audio reference mixer — mixdown-equivalence (2026-06-13)**:
  `wai-rs` `audio_scene` feature (reuses the world `Fx` floor + blake3,
  no codec/ML deps): the `WSCN` container (scene contract + object table
  + asset table + optional seal, mirrors WWLD) and a fully deterministic
  mixdown — `Fx` position interpolation, `inverse_clamped` distance
  attenuation, `sqrt_constpower` pan (sqrt only, no transcendental
  tables), `Fx` accumulate, round-half-away i16 quantization in integer
  space. `mixdown_hash` (BLAKE3 over canonical i16LE PCM for a declared
  layout) is the **mixdown-equivalence** quantity; `scene_hash` is the
  identity. 5 tests prove it: centered = equal L/R, hard-left = silent
  right, deterministic + layout-bound hash, distance attenuates, a
  moving source sweeps L→R + container round-trips. Sources arrive as
  decoded mono PCM keyed by content hash — decoding the `wai.audio.*`
  assets is the sink's job. clippy clean. Next: conformance corpus
  bin + Web Audio browser sink + demo.
- **Content taxonomy + cargo roadmap; spatial-audio Phase 1 begun (2026-06-13)**:
  mapped the full space of web-transported content WAI's approach covers
  (`research/content-taxonomy.md`) — a capture→synthesizable→executable→
  symbolic spectrum × four levers (codec dispatch, neural prior,
  instructions-at-sink, determinism+receipts), with a per-row
  mechanism/extension/status table and the white space called out. A
  phased build plan (`research/cargo-roadmap.md`) takes each white-space
  class through the proven interactive-worlds recipe. **Phase 1 kicked
  off — object-based spatial audio** (`wai.audio.scene`): the audio
  analog of interactive-worlds (ship sound objects + a scene, render to
  the sink's layout; **mixdown-equivalence** = replay-equivalence with
  the tick replaced by the audio sample). Spec drafted
  (`research/spatial-audio-draft.md`, `WSCN` container, `Fx` mixdown,
  content-addressed `wai.audio.*` sources, stereo bit-exact floor +
  binaural perceptual tier, WAI2 recorded-mix fallback, JWP receipt
  profile); `wai.audio.scene` registered + tiered Lossless. 94 tests
  green. Remaining phases: haptics, live state-feeds, digital humans,
  volumetric/4D splats.
- **Real SPZ decode — `wai.splat.spz` end to end (2026-06-13)**: the
  splat gallery now renders a genuine SPZ [name removed 2026-09-30] file, not a
  mesh-derived stand-in. New `wai-spz-gen` crate depends on the field's
  `spz` reference library (the sink-supplied codec WAI dispatches to),
  encodes a 2,400-Gaussian cube cloud to a real `.spz` (validated by
  re-loading through the library), and wraps it as a content-addressed
  `wai.splat.spz` WAI object (`nebula.spz.wai`). `gen-gallery` adds it
  to the scene's asset table (role 1) beside the glb. Browser:
  `WorldSink.asset_payload(name, env, cap)` generalizes the §4.1 verify
  to any capability; the page gunzips (DecompressionStream) and decodes
  the SPZ bitstream in-tab — 16-byte header, non-interleaved arrays,
  24-bit fixed-point positions, log scales, smallest-three quaternions,
  byte/255 opacity+color — into per-Gaussian world covariance for the
  EWA pipeline. Decoded count matches the encoded 2,400 (validates the
  layout); the cube silhouette + surface-aligned discs validate the
  quaternion math. Page copy corrected: the splat view decodes a real
  SPZ asset (no longer "derived from mesh"). Verified locally (2,400
  Gaussians, zero errors); prod serves the identical artifacts.
- **Anisotropic EWA splatting (2026-06-13)**: upgraded the splat
  renderer from isotropic billboards to real 3D Gaussian splatting. Each
  splat carries an anisotropic 3×3 covariance (surface-aligned flat
  discs: wide in the triangle tangent plane, thin along the normal); the
  vertex shader projects it to a screen-space conic
  (Σ' = J·W·Σ·Wᵀ·Jᵀ with the perspective Jacobian), emits a tight
  3σ bounding quad, and the fragment evaluates the conic — the textbook
  3DGS pipeline, still CPU-sorted back-to-front and alpha-blended.
  Cam uniform carries view + proj separately (+ viewport/focal) for the
  Jacobian. Visibly sharper, surface-coherent splats; proven on prod,
  zero errors.
- **3D render layer — PBR mesh + Gaussian splats (2026-06-13)**: the
  `/worlds` gallery now renders the content-addressed, hash-verified
  `wai.mesh.glb` two ways in a real 3D path (perspective camera + depth
  buffer, all in `wai-web/demo/world.js`, no wasm change): a **solid
  PBR-shaded cube** (flat normals from screen-space derivatives, Lambert
  + spec + sky-ambient, gamma) and a **Gaussian-splat cloud** —
  instanced camera-facing quads with a Gaussian falloff, back-to-front
  depth-sorted each frame (painter's order) and alpha-blended. The splat
  cloud is *derived from the same verified mesh*, making the thesis
  visible: one content-addressed asset, two sink-side renderings,
  presentation never touches the state hash. SPZ binary decode stays
  sink-supplied (stated honestly on the page); the renderer is the real
  deliverable. Two gallery vectors (`gallery` / `gallery_splat`); the 2D
  sim path is untouched. Proven on prod: both render, 768 splats from
  the cube, zero errors.
- **Receipt chaining + work-meter + `wai.splat.spz` (2026-06-12)**:
  three follow-ons. (1) `SessionReceipt` gains `parent_receipt_hash`
  (BLAKE3 of the prior receipt's canonical bytes — the JWP linkage) +
  `receipt_hash()` + `verify_chain`: successive sealed sessions of one
  world (shared `world_hash`, `group_id`+1) form an append-only signed
  timeline. (2) An **honest energy meter**: receipts carry exact,
  signed, deterministic `work_steps` (Σ node-steps, from
  `collect_checkpoints_and_work`) AND a separately *carried*
  `joules_micro` — never derived in the receipt; the browser supplies a
  wall-clock estimate at a **declared** coefficient (labeled "est@4W"),
  native sinks supply real RAPL, an unmetered sink signs `0`. (3)
  Registered **`wai.splat.spz`** — Gaussian-splat scenes as an asset
  capability parallel to `wai.mesh.glb` (presentation only, never in the
  state hash; the win is content-addressing + dedup, not payload size;
  renderer sink-supplied). Browser `seal_session(seed, id, joules,
  group, parent)` chains via localStorage; the `/worlds` HUD shows
  `chain #N ↳linked · N steps · ≈X mJ est@4W · ✓ re-verified`. Proven on
  prod: two seals chain (g0 parent null → g1 parent = g0.receipt_hash),
  exact work 248 steps, est joules carried, zero errors. Spec §8 updated
  (chaining + the two-quantity energy model). 94 native tests, 10/10
  vectors, clippy clean.
- **Asset path SHIPPED — `wai.mesh.glb`, content-addressed (2026-06-12,
  extension §4.1)**: worlds are made of WAI objects. New
  `world::asset_content_hash` (BLAKE3 `wai:asset\x01` over a complete
  WAI envelope) + `DOMAIN_ASSET`; native tests PROVE the load-bearing
  claim — a scene with an asset table has the byte-identical state hash
  AND world hash as the same scene without it (assets are
  presentation-only, never in the state hash), and the asset hash binds
  the exact envelope. Generator `gen-gallery` writes a `gallery.wwld`
  scene + `assets/cube.glb.wai`: a real glTF 2.0 Binary cube
  (hand-written GLB writer) wrapped as a `wai.mesh.glb` WAI envelope and
  content-addressed in the table. Browser: `WorldSink.assets()` +
  `asset_glb(name, envelope)` (verifies content hash in WASM, rejects a
  mismatch, unwraps the WAI envelope to the `.glb` payload); the page
  decodes the GLB (minimal JS decoder) and renders the verified mesh as
  a spinning wireframe. Proven on prod: `1/1 verified ✓ · cube.glb`,
  and the state hash stays fixed while the mesh spins — §4.1 made
  visible. 92 native tests, 10/10 vectors, clippy clean.
- **Session receipts + seal-to-replay SHIPPED** (2026-06-12, extension
  §8 implemented): `world/receipt.rs` builds a `SessionReceipt` — a JWP
  *profile*, not a fork: op batches (one per checkpoint interval,
  content-hashed `wai:world-ops\x01`) are Merkle-bound with JWP's exact
  leaf/node domain separators, and one Ed25519 signature covers the
  world-profile payload (world hash + checkpoint `state_hash` list +
  carried microjoules). Joules are **carried, never fabricated** — the
  builder takes the figure; an unmetered session signs `0` and says so.
  New `world_hash` (BLAKE3 over scene = contract + tick-0 graph, the
  §7.4 shareable identity), `collect_checkpoints` (unsealed replay →
  checkpoint list). Browser: `WorldSink.seal_session(seed, signer_id,
  joules)` → `SealedSession{replay_bytes, receipt_json, root_hash, …}`;
  the `/worlds` page's "⏷ seal & download replay" turns a live, shared,
  edited world into a portable `.wwld` + `.receipt.json`, re-opens the
  fresh replay in a throwaway sink to prove it re-verifies, and shows
  the Merkle root + batch count in the HUD. Proven on prod: fracture a
  tower across the relay → seal → 3-batch receipt, valid sig, replay
  re-verified. 90 native tests, 10/10 vectors, clippy clean.
- **Cross-device multiplayer SHIPPED** (2026-06-12, `wai-relay/` +
  `wai-web` Actor binding): every tab/device is a signed actor. New
  `Actor` wasm-bindgen class (browser-held Ed25519 key; clicks author
  real `CausalOp`s; Lamport discipline via `observe_lamport`);
  `wai-world-relay` reference transport (axum WS, per-room fan-out,
  retains the signed op log for late joiners, **verifies every op's
  Ed25519 signature at the door** with the same `wai` crate — it
  adjudicates nothing; sinks converge because the math does). Page
  transport is relay-first (room in the URL hash) with BroadcastChannel
  fallback. Deployed at `wai.transaction.science/worlds-ws/*`
  (launchd `com.wai.world-relay`, localhost:9670). Proven on prod with
  two ISOLATED browser contexts: point-blank axial shoves snap two
  members (23→21 nodes) and both contexts converge — "✓ hashes agree".
  Also fixed `build.rs` emitting libFLAC link flags regardless of
  features (pkg-config probe side effect): now gated on `codecs`.
- **Browser sink SHIPPED** (2026-06-12, `wai-web/src/world.rs` +
  `wai-web/demo/world.{html,js}`): a `wai.world.replay` sink in a tab.
  NOT a port — the `wai` crate itself now compiles to wasm32 via the
  new default-on **`codecs` feature** (C codec chain gated; envelope +
  `world` engine build with `--no-default-features --features world`),
  and wai-web binds it (`WorldSink`: advance/particles/members/
  state-hash/verify-seal; sim state stays exact Q32.32, f32 only for
  display). WebGPU renderer (instanced SDF circles + member lines).
  Proven in a real browser [name removed 2026-09-30] via Playwright: the 1,282-byte truss_fracture
  payload replays 120 ticks, the member snaps, and the browser's
  BLAKE3 state hash equals the native corpus hash byte-for-byte —
  **replay-equivalence across arm64-native and wasm32 in a browser [name removed 2026-09-30],
  demonstrated**, with the Ed25519 seal verified in-page.
- **Solver-backed registrant `wai.world.sim.truss0`** (2026-06-12,
  `world/truss.rs`): structural simulation — particles + distance
  members solved per tick by position-based dynamics (fixed 8-pass
  Gauss–Seidel projection in canonical member order), entirely in
  `wai.det.fixed64` including a new floor-rounded `Fx::sqrt` (128-bit
  integer Newton). Members carry a yield strain measured on predicted
  positions before the solve; an overloaded member **fractures** as an
  emergent solver consequence — two actors pulling the same member in
  opposite directions (the SUM rule feeding the solver) snap it
  bit-identically on every machine (`truss_fracture` corpus vector).
  Registration constants: `G = −10`, `ITERATIONS = 8`, ground at
  `y = 0`, velocity damping `63/64`/tick (static equilibrium).
  `LiveWorld` gained `state()`/`current_tick()` accessors (what a
  renderer reads). Where the float-based CAD solvers can't transplant
  (floats aren't cross-platform bit-exact), solver *form* does:
  iterative constraint projection in the deterministic profile.

### Inter-frame video, categorical/VQ images, and real-model validation

- **`wai.neural.dcvc_rt` / `wai.neural.dcvc_fm` — inter-frame video decoder
  implemented** (`neural/video_dcvc.rs`). "Requires a GPU" [name removed 2026-09-30] was a *speed* claim:
  inter-frame decode is a frame loop that propagates a reference feature, and
  ONNX Runtime runs it on CPU. The per-frame entropy decode reuses the
  CompressAI-validated hyperprior path; the temporal net is sink-supplied.
  Validated end-to-end (3 inter-frames, feature propagation confirmed).
- **`wai.neural.glc` — categorical / VQ decoder implemented** (`neural/glc.rs`).
  Not the Gaussian-conditional family: codebook indices decoded via a
  per-position softmax → `pmf_to_quantized_cdf` (bit-exact with CompressAI's
  C++, unit-tested) → a bypass-free categorical rANS over a checkerboard → VQ
  synthesis. Surfaced that categorical entropy is the most determinism-sensitive
  path — a per-position CDF flips on ~1e-5 logit drift where Gaussian scale
  buckets don't.
- **Real-model validation** beyond synthetic stand-ins. SNAC's genuine
  `snac_24khz` now exports (`tools/wai_snac_export.py`: zero-noise so the decode
  is deterministic at all + a reshape-free snake) and `wai-rs` decodes it to
  ~9e-7 of the torch reference. A real PTQ-trained, activation-calibrated
  `int_synth` model decodes **bit-identically** (max|Δ|=0) to its integer-forward
  reference (`tools/wai_int_synth_train.py`; also fixed a zero-gradient
  `torch.floor` bug in the QAT trainer).

### Learned-image codec family — decoders implemented and validated

- The reference crate now decodes the full conditional-entropy family
  natively, not just the factorized baseline: **`wai.neural.mbt2018_mean`**
  (mean-scale hyperprior, Gaussian-conditional), **`wai.neural.elic`**
  (checkerboard spatial context), **`wai.neural.mlicpp`**
  (channel-autoregressive multi-reference), and the integer-exact
  **`wai.neural.int_synth`** (all-integer transposed-conv synthesis — a
  decode that is bit-identical on every machine *by construction*, the
  first capability that can soundly back `intent=replicate` cross-hardware).
  The rANS + entropy decode live in `wai-rs`; the analysis/synthesis nets
  are sink-supplied ONNX.
- **Current-SOTA entries added as live routes:** `wai.neural.mlicv2`
  (MLICv2/v2+, the current SOTA learned image) and `wai.neural.cmic`
  (Content-Aware Mamba) both decode through the multi-reference core —
  their entropy model is the same channel-slice + checkerboard structure;
  the gains live in their sink-supplied transforms.
- **Validation against the reference codec.** `bmshj2018` and
  `mbt2018_mean` are checked byte-for-byte against pretrained CompressAI
  models (reconstruction matches to within final-layer float rounding);
  `elic`/`mlicpp` are checked end-to-end against minimal CompressAI-
  compatible models; `int_synth` against a constructed integer model.
  Surfacing this also caught (and the spec now documents) a real
  determinism hazard: a learned codec's entropy decode desyncs unless the
  encoder and decoder compute the scale buckets with the *same* model —
  float drift between two implementations of `h_s` flips bucket
  assignments and corrupts the stream.

### Audio: SNAC hierarchical codec

- **`wai.neural.snac`** — SNAC's multi-resolution RVQ, distinct from the
  flat single-grid RVQ runtime: N variable-length code streams at
  hierarchical rates, decoded via dynamic-arity ONNX inputs. Ubiquitous in
  low-bitrate TTS.

### Removed: off-charter Python runtime (`wai-neural/`)

- Moved `wai-neural/` OUT of this repo to `../../wai-neural-prototype/`.
  It was a parallel **Python** runtime (`WAI4` generative-semantic
  container + CLIP-embedding payloads + SD-XL-Turbo regeneration +
  CLIP-similarity conformance) — a disconnected second implementation in
  the wrong language. The reference implementation is **Rust** (`wai-rs`);
  nothing referenced the Python package. The *idea* it proved — a
  behavioral/perceptual neural path that conforms by semantic similarity
  rather than a weights hash — is sound and belongs in `wai-rs` as a future
  Rust path alongside the bit-exact, prior-pinned path. The Python is kept
  only as a prototype reference (and it hardcodes a machine-local model
  path, so it never ran portably). This does NOT touch the Python
  model-export tools in `tools/`, which are legitimate dev tooling.

### Neural condition: working end-to-end

- `tools/wai_encodec_encode.py` / `wai_encodec_decode.py` — first
  working **neural capability** (`wai.neural.encodec32`): source
  encodes audio via locally-installed EnCodec, ships only the
  quantized tokens in a standard WAI envelope (236 MB of model
  weights stay on the source's disk, never enter the envelope); sink
  with EnCodec installed decodes. Measured: **5 s of audio → 2.3 KB
  envelope (141× compressed)**, model on the wire = 0 B.
- Rust SDK gains `wai inspect` CLI: parses an envelope, resolves the
  capability + fallback against `sink_capabilities()`, and prints
  verdict (`DECODABLE via primary` / `via fallback` / `INERT here`).

### Browser runtime (`wai-web/`)

- New crate: wasm-bindgen wrapper exposing `parse(bytes) → WaiEnvelope`,
  `pack(manifest_json, payload)`, and `known_capabilities()`. 107 KB
  wasm + 16 KB JS glue.
- `wai-web/demo/wai.js` — thin JS shim with `WaiRuntime.register(cap,
  handler)` dispatch. Backends are **pluggable** (browser-native APIs,
  transformers.js, WebGPU compute, WebNN, ONNX runtime web) — WAI
  doesn't bundle any. The sink-side compute fabric is heterogeneous;
  the deployer picks per capability.
- `wai-web/demo/index.html` — drop-and-decode demo with handlers wired
  to: `createImageBitmap` for `wai.image.{png,jpeg,avif,jxl}`,
  `AudioContext` for `wai.audio.{opus,flac}`, placeholders for
  `wai.neural.*` showing where transformers.js / WebNN / WebGPU plug
  in. Run with `python3 wai-web/demo/serve.py`.
- 4 sample `.wai` files in `wai-web/demo/samples/` covering PNG, JXL,
  Opus, and EnCodec for testing.

### Spec: v1.1 multi-rendition envelope (SPEC.md §6.1)

- New magic `WAI2` for envelopes that carry multiple renditions of the
  same content (e.g. EnCodec + Opus, or AVIF + JPEG-XL). Reframed from
  "graceful fallback for missing capability" to **policy selection** —
  in a self-contained ecosystem every sink has every required
  capability; renditions exist so the sink picks by compute budget,
  bandwidth, or runtime preference (browser-native vs WebGPU vs WebNN
  vs native ML). v1.0 (`WAI1`) remains valid and unchanged.

## v0.3.0

**Pivot**: WAI is the **envelope + capability-dispatch + neural-shared-
prior** standard. The zeroth condition is a *menu of registered SOTA
codec capabilities* (AVIF, JPEG-XL, PNG, JPEG, Opus, FLAC, AV1, zstd,
XZ), not a custom codec. The Python prototype and all hand-rolled
codec implementations from v0.2.x were removed; Rust is the single
reference implementation and the SDK.

### Standard

- `SPEC.md` rewritten as a capability-menu spec. Mandatory floor:
  `wai.image.png`, `wai.audio.flac`, `wai.text.zstd` (every conforming
  sink MUST decode these). Recommended modern set: `wai.image.jxl`,
  `wai.image.avif`, `wai.image.jpeg`, `wai.audio.opus`, `wai.video.av1`,
  `wai.video.av1.lossless`, `wai.text.xz`. Neural capabilities
  (`wai.neural.*`) remain OPTIONAL and sink-advertised.
- Container envelope (`WAI1` magic + length-prefixed manifest + length-
  prefixed payload) and manifest schema are unchanged from v0.2.x.
- Conformance: bit-exact decode-equivalence with each capability's
  reference library (libjxl / libavif / libopus / libFLAC / rav1e / etc.).

### Rust reference (`wai-rs/`, crate `wai` 0.3.0)

- `cargo build --release` produces `libwai.dylib`/`.so`/`.dll` (cdylib),
  `libwai.a` (staticlib), and `libwai.rlib` (Rust-native). The crate
  type list is `["lib", "cdylib", "staticlib"]` — drop-in for any
  language's C-FFI binding.
- Codec wrappers: `image` (PNG, JPEG via `jpeg-encoder`, AVIF via
  `ravif` + `libavif-image` for decode, JPEG-XL via `jpegxl-rs`), `opus`
  (RFC 7845 Ogg-Opus container so output drops into ffmpeg/VLC/browsers),
  `claxon`+`flac-bound` for FLAC, `rav1e`+`dav1d` for AV1 (lossy +
  lossless), `zstd`+`xz2` for text.
- Generated C header: `include/wai.h` is regenerated by `build.rs` via
  `cbindgen` on every build. SDK consumers get a typed header rather
  than hand-writing `extern "C"` decls.
- Cross-platform linkage: `build.rs` probes `pkg-config` for libflac;
  the macOS-hardcoded `.cargo/config.toml` from v0.2.x is gone. Builds
  clean on macOS (Homebrew) and Linux (apt) without `RUSTFLAGS`.

### C FFI / SDK surface

All `wai_*` functions return a `wai_status_t`:

| code | meaning |
|---|---|
| `WAI_OK` (0) | success |
| `WAI_ERR_INVALID_ARG` (1) | null pointer / mismatched length / out-of-range dimension |
| `WAI_ERR_CODEC` (2) | codec-level failure (details in `wai_last_error()`) |
| `WAI_ERR_CONTAINER` (3) | bad magic / truncated envelope / unparseable manifest JSON |
| `WAI_ERR_PANIC` (4) | internal Rust panic caught at the boundary (a wrapper bug — never UB into the C caller) |

Every FFI function is wrapped in `std::panic::catch_unwind` so Rust
panics never unwind into C (UB). `av1_decode` validates header bounds
to prevent libdav1d's C-level `abort()`. `wai_last_error()` is per-
thread, populated on any non-OK return.

### Testing

- **13 lib tests**: per-codec round-trip + envelope round-trip
- **21 FFI robustness tests**: null pointers, length mismatches,
  dimension overflow, malformed magic, truncated payloads, bad JSON,
  invalid Opus sample rates, zero fps, garbage decoder input
- **10 fuzz smoke targets**: 1024 adversarial random byte sequences per
  decoder asserting no panic. Stable Rust, no `cargo-fuzz` toolchain
  needed. Found and fixed a real envelope-truncation panic.
- **`wai_bench` corpus walker**: quick mode (4 images + 4 audio + 3
  text + 2 video) for spot-checks; `--full` walks the entire corpus
  (169 images + 5 audio + 13 text + 12 video sequences = 588 codec
  runs in ~9.4 min). Per-codec aggregate stats: runs, ok / fail counts,
  WAI vs reference byte sizes, PSNR delta vs ffmpeg, encode time.
- **`--report file.json`** writes a structured bench report;
  `wai_bench_diff baseline.json head.json` flags regressions on bytes
  (±5% default tolerance), encode time (±25%), or new failures. Non-zero
  exit on regression — ready for CI gating.
- **Hosted CI** [name removed 2026-09-30] (`.github/workflows/ci.yml`): ubuntu + macOS
  matrix, apt/brew codec deps, `cargo test --release`, cdylib/staticlib
  artifact verification, quick bench, JSON report uploaded as artifact.

### Honest codec measurements (vs ffmpeg reference, full Kodak/Tecnick/
CLIC/synthetic corpus = 169 images, plus SQAM audio + Silesia text + Derf CIF video)

| codec | runs | size vs ref | PSNR Δ vs ref | notes |
|---|---:|---:|---:|---|
| PNG | 169 | — | lossless | ffmpeg-interop ok |
| JPEG q=50 | 169 | +12.7% | +0.27 dB | q-knob calibrated to libjpeg |
| JPEG q=80 | 169 | +19.7% | −0.31 dB | residual gap is pure-Rust vs libjpeg-turbo asm |
| AVIF q=70 | 169 | — | — | ffmpeg-interop ok |
| JPEG-XL lossless | 169 | — | lossless | bit-exact via libjxl |
| FLAC | 5 | **−10.1%** | lossless | WAI compression_level 8 vs ffmpeg default 5 |
| Opus 64k | 5 | — | — | standard Ogg-Opus container, ffmpeg-interop ok |
| zstd-19 | 13 | ±0.0% | — | **byte-identical to system zstd** |
| xz-6 | 13 | ±0.0% | — | **byte-identical to system xz** |
| AV1 lossy | 12 | +28.3% | — | rav1e vs ffmpeg libsvtav1 (different encoders) |

All 588 corpus runs and all 44 tests pass.

### Removed

- The Python prototype (`wai/`) and Python BD-rate harness (`bench/`).
  v0.2.x's hand-rolled wire format is dead; Rust is the single source
  of truth.
- The `webm_lottie_new/wai-rs/` pre-pivot Rust crate (Wasm-decoder-in-
  bundle model). Replaced by the clean `wai-rs/` SOTA-wrapper crate.

---

## v0.2.0 — superseded (historical record)

The v0.2.x direction was hand-rolling custom codec wire formats (DCT +
quantization + rANS for image, MDCT for audio, I/P inter-frame for
video). Two byte-interoperable reference implementations (Python +
Rust) were built and cross-validated.

**Why superseded**: re-implementing JPEG/MPEG/Opus poorly is strictly
worse than wrapping the field's SOTA libraries. WAI's actual value is
the envelope + capability dispatch + the neural-shared-prior model,
which v0.3 keeps. Custom codecs are gone; standard libraries are in.

### What v0.2 had

- `SPEC.md` defined byte-precise wire formats for image (8×8 DCT-II +
  YCbCr + JPEG-Q + zigzag + DC-Δ + rANS), audio (orthonormal MDCT +
  sine window + TDAC), video (I/P inter-frame with integer-pel motion).
- Python (`wai/`) + Rust (`wai-rs/`) implementations, byte-equivalent on
  the lossless paths (rANS, container) and reconstruction-equivalent on
  the lossy paths.
- Honest measured standing: image at JPEG-1992 parity (~5% behind
  libjpeg), audio uniform-MDCT (not perceptually competitive with MP3),
  video I/P at ~41% smaller than MJPEG floor but far behind H.264.

### Licensing (unchanged)

- Apache-2.0 for the standard and the reference implementation.
  Royalty-free implementation is a precondition for adoption (VP9/AV1
  open-codec precedent; the H.264/H.265 royalty trap is avoided).