wai: wai generate and wai score, and the instruction object of wai.generate.spectral (form v0) #197

Merged

dcharlot merged 12 commits from wai/generate-score-cli into main 2026-10-10 22:59:43 -04:00

Conversation 0 Commits 12 Files changed 25 +2265 -24

2026-10-10

Transaction Science

Merge origin/main into wai/generate-score-cli ...

Brings in #200 (HTTP receipts and keys) and #206 (wai-ops) among the rest
of main. One conflict, wai/CHANGELOG.md: both sides added entries at the
top of v0.4.0. Kept every entry; this branch's "Generated pictures and
scores" stays on top as the newest, main's entries follow unchanged.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

815f3fb8c2

• View at this point in history

2026-10-10 17:18:45 -04:00

2026-10-09

Transaction Science

wai-web wasm32 CI: the browser generator and score synth make every native pin ...

The CHANGELOG said every pin equals the shared browser binding's output, and the
browser sink's form-v0 reading has no CI of its own, yet wai-web-wasm.yml ran no
generate or score check. demo/verify_generate_wasm.mjs runs generateSpectral at
every size the generate vectors pin (sizes.json, instr-v0.json's four listed sizes,
and the first listed size of each object at a bound of form v0), checks each
object names the committed model, and runs ScoreSink over every score-conformance
vector for its pcm_sha256, mixdown hash and score hash. The workflow runs it on the
node build and now also triggers on wai/score-conformance/. 35 checks pass locally
on a wasm32 build of this tree.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

4c09bea76a

• View at this point in history

2026-10-09 17:17:06 -04:00

Transaction Science

wai REGISTRY: the status note names wai.generate.spectral among the registrations v0.4.0 completes ...

The hand-written Status note listed the integer codecs whose registrations v0.4.0
completes and left out wai.generate.spectral, which SPEC §8 now says §5 completes.
The note sits outside the generated region the library test compares.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

cdc351b33c

• View at this point in history

2026-10-09 17:11:00 -04:00

Transaction Science

wai generate.spectral: the model is the parameter set, so the envelope carries no pin; inspect and unwrap report the payload ...

Form v0 names the model inside the payload by the SHA-256 of one file, and SPEC §5
never said how that relates to §3.1's parameter-set pin, which governs every
GenerativeDeterministic capability: whether a generate envelope carries a pin, or
which wins when a pin and `model` disagree. The library registers the set as
several files, so a form-less pin over the very file `model` names would be
refused form_required. SPEC §5 now states the relation: the model file is the
parameter set and `model` its identity; an envelope that names the generator
carries no `prior` and a sink refuses one that does, whatever it names. §3.1 says
so beside the pin rules, the non-conforming list names it, and prior-carriage §7's
row says the set is never pinned and names the file. No library source changes.

`wai inspect` and `wai unwrap` reported a generate envelope with a corrupt
payload as conforming ("component roles ✓", no payload line). They now report
the payload (form v0, with the reader's reason), the intent and any pin of every
envelope or WAI2 component that names the generator, ✓ or ✗; inspect's verdict
says the envelope does not conform. inspect_and_unwrap_report_the_generators_envelope
covers a prompt byte changed in place, the intent improve, a pin over the model
file itself, and a WAI2 component with and without a pin.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

7e3b883074

• View at this point in history

2026-10-09 17:08:26 -04:00

Transaction Science

wai generate.spectral: the generator's envelope carries intent create, and wrap writes no other ...

SPEC §5 said the envelope's intent is `create`, but `wai wrap` refused only
`replicate` and wrote wai.generate.spectral envelopes with `--intent improve` or
`--intent banana`, and the non-conforming list did not mention intent. The
sentence is now a rule: an envelope that names the generator, as capability or
fallback, carries `create`, one with another intent is non-conforming, and wrap
refuses any other. wrap_writes_the_generator_only_with_intent_create covers
improve, an open string, a case variant, the empty string and the fallback.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

da03ef1524

• View at this point in history

2026-10-09 17:02:55 -04:00

Transaction Science

wai generate.spectral form v0: accepted objects at every bound, and each refusal's reason ...

instr-v0.json had one accepted object and 28 refused ones, none at an inclusive
bound, so a second sink stricter by one passed every vector; and the CLI test held
only that each object was refused, not why, so a rule relaxed (a quotation mark let
into the prompt) stayed green while the canonical-bytes check refused the object
in its place.

- The reader (wai_cli/instr.rs) refuses with a reason code, the first rule broken
  in a stated order; a member given twice is refused as such (strict_json).
- The vector gains twelve accepted objects (seed 0 and 2^32 - 1, steps 1 and 64,
  an empty prompt, 256 bytes, 0x20 and 0x7e, JSON's structural characters, a side of
  1, a side of 4096 and 2^23 pixels, 16 sizes, the w-then-h order), each with the
  SHA-256 of its first listed size made natively; eight more refused objects (types,
  a fraction in a size, a negative side, an escaped line feed, a backslash, cut-short
  JSON); and a reason for every refused object.
- tests/cli_generate_score.rs asserts each refusal's reason through `wai generate
  --instr` and `wai wrap`, and makes and wraps every accepted object.
- tests/generate.rs states form v0 a second time, independently of the reader, and
  holds the object, the accepted and the refused lists to it, making each accepted
  object's output with the library.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

f0107b07b8

• View at this point in history

2026-10-09 17:01:04 -04:00

Transaction Science

wai score build: refuse a member given twice in the authoring form ...

The authoring form (score §2.3) is the definition, yet a note holding both
"start":0 and "start":5 built with start 5 (serde_json keeps the last value) and
the same hash as a note with start 5 alone; another JSON reader could keep the
first. wai_cli/strict_json.rs reads JSON with every object's members distinct and
refuses a repeat, and build reads its document through it. Two refused cases are
added: a note field and a document member, each given twice.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

95fa7244f6

• View at this point in history

2026-10-09 16:56:41 -04:00

Transaction Science

wai score render: refuse a contract that is not one channel of wai.det.fixed64 ...

render printed "channels":2 beside a mono mixdown of `duration` samples when a WSCR's
contract said two channels, and rendered a contract whose numeric profile was
wai.det.float64; a verifier following SPEC §7 Digests would read that PCM as stereo
frames. render now refuses both, as build refuses to write them.
score_render_refuses_a_contract_it_does_not_render patches the committed scale.wscr
in place for each case.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

887c99a61f

• View at this point in history

2026-10-09 16:54:46 -04:00

Transaction Science

wai cli tests: the note-count guard is what refuses the crafted WSCR ...

score_render_refuses_a_note_count_its_bytes_cannot_hold wrote a file with no
contract section, so the reader refused it with MissingSection(0x01) before
parse_notes reserved anything, guard or no guard: the test checked only which
message came first. The file now carries a well-formed contract (section 0x01) and
a notes section of only the count (0x02), and the same file declaring no notes is
shown to render, so without the guard the reader reaches Vec::with_capacity.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

0d123ab7f2

• View at this point in history

2026-10-09 16:53:08 -04:00

Transaction Science

wai cli tests: the two instr_sizes callers each write their own object file ...

instr_makes_the_listed_small_sizes and instr_makes_the_listed_large_sizes run at
once in the release step, and both removed and rewrote one scratch file,
object.instr, while the other's `wai generate --instr` children read it. A remove
or a truncating write landing as a child opened the file would refuse the object
and fail the test. Each caller now names its own file.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

35debe1706

• View at this point in history

2026-10-09 16:52:15 -04:00

Transaction Science

wai SPEC §5: the payload of wai.generate.spectral is an instruction object (form v0) ...

The registration is completed as §8 permits: the row named the generator and no byte
layout. The payload is 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, each within stated
bounds and with one serialization per object, so an envelope's content hash names the
instructions. A sink makes only listed sizes, with that model only; conformance is
generation-equivalence at each listed size, and a sink holding a size's SHA-256 does
not present a picture that differs or accept one on a 64-bit FNV-1a alone. The
envelope's intent is `create`: a generator does not back `replicate` (§7), and this
keeps that rule.

The registry row names the payload; §8 lists the completion. score §2.3 documents the
integer authoring form `wai score build` reads, and §4 the pcm_sha256 every vector now
states. CI builds the `wai` binary with `generate` and `score` in both profiles of the
integer-kernel steps and runs tests/cli_generate_score.rs there; the two larger listed
sizes of instr-v0.json run in the release step.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

3770d1ab8e

• View at this point in history

2026-10-09 10:20:14 -04:00

Transaction Science

wai: wai generate and wai score build|render, with native pins a sink checks against ...

`wai generate` makes a wai.generate.spectral picture with the reference generator at
any size and prints its SHA-256; with --instr it makes a listed size of an instruction
object of form v0, which it reads as a sink does, and --write-instr writes one.
`wai score build` writes a WSCR from integer note fields only; `wai score render`
prints samples, pcm_sha256, mixdown_blake3 and score_hash, and --pcm writes the samples.

Vectors: tests/vectors/generate/sizes.json (256x128, 640x320) and instr-v0.json (the
object, its four listed sizes and 28 objects a sink refuses); score-conformance gains
pcm_sha256 for every vector and notes/<name>.json, the authoring form, which
`wai score build` turns back into the committed WSCR byte for byte. Every pin was made
natively and equals the shared browser binding's output.

`wai wrap` refuses a wai.generate.spectral payload that is not a form-v0 object, in
every build. No library source changes.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

a220d16a48

• View at this point in history

2026-10-09 10:01:31 -04:00