wai: register the byte-exact JPEG-AI decoder + restore the byte-equality matrix #50

Merged

dcharlot merged 6 commits from wai-jpegai-integrate into main 2026-08-01 05:43:10 -04:00

Conversation 0 Commits 6 Files changed 14 +586 -19

2026-08-01

Transaction Science

wai: document the reconstruction binding + dispatch/conformance changes ...

- SPEC §5: reference extensions/reconstruction-binding.md alongside the other
  provenance sub-specs, so the new capability is discoverable from the spec.
- CHANGELOG: entries for the reproducible-reconstruction binding (the C2PA/
  JPEG-Trust hard binding over decoded pixels), the wai.image.jpegai dispatch
  now preferring the byte-exact decoder, and the restored byte-equality matrix.

Docs only; no code change.

Co-Authored-By: Claude Opus 4.8 <[email protected]>

afb4e603cc

• View at this point in history

2026-08-01 05:25:41 -04:00

Transaction Science

wai-rs: mark the legacy ORT mANS decoder, point new code at the canonical one ...

Two mANS decoders exist: the crate-root `jpegai_mans` (byte-exact, in the cross-arch
byte-equality matrix, used by the deterministic decode) and this `neural/jpegai_mans`
(the older port backing the float ORT fallback). They are cleanly separated by path —
only `neural/jpegai.rs` uses this one — so rather than force a risky merge of the
fallback's entropy stage, mark this copy legacy/ORT-only with a pointer to the
canonical crate-root decoder and the deprecation direction: it stays only while the
ORT fallback does, and goes when the deterministic path fully supersedes it.

Also verified this pass (not a code change): the byte-exact decode compiles to
wasm32-unknown-unknown under `neural_int`, so the deterministic reconstruction runs
on every host including the browser — the portability half of the byte-exact contract,
alongside the entropy stage's presence in the cross-arch byte-equality CI matrix.

Co-Authored-By: Claude Opus 4.8 <[email protected]>

5ac0d5c86c

• View at this point in history

2026-08-01 01:24:54 -04:00

Transaction Science

wai: reproducible-reconstruction binding — a C2PA/JPEG-Trust hard binding over pixels ...

The byte-exact decoder's payoff: a provenance binding neural-coded media normally
cannot carry. A hard binding hashes asset bytes; over the decoded image it was
impossible, because a neural decode isn't reproducible — JPEG AI's own normative
conformance point is the entropy-decoder output (not pixels), and its multi-branch
design lets one codestream reconstruct three ways, so a decoded-pixel hash isn't
portable. Provenance over the rendered result has had to fall back to fuzzy soft
bindings.

WAI decodes to RGB deterministically, which restores a stable, portable decoded-pixel
hash — pinning one reconstruction is a determinism guarantee beyond what the base
standard requires. This adds:

- reconstruction_binding (feature provenance): ReconstructionBinding::bind hashes the
  canonical RGB8 buffer (row-major, no padding) under a descriptor that pins the
  reproducible decode (synthesis branch / profile / level / layout / color space);
  verify re-decodes-and-compares. SHA-256 — the C2PA default — so the digest is
  reproducible by any validator. bind/verify/tamper/JSON tests pass.
- extensions/reconstruction-binding.md: the standard. Decode-equivalence conformance;
  carriage as a custom C2PA/JPEG-Trust assertion (science.transaction.wai.hash.
  reconstruction, CBOR, beside the standard byte-binding); and a note that a signed
  binding over reconstructed content is a substrate for machine-readable AI-content
  marking (e.g. EU AI Act Article 50, from 2 Aug 2026) — stated as a capability, not a
  legal claim.

It is the reconstruct-side analogue of the provenance extension: that binds the opaque
carried bytes; this binds the reconstructed pixels.

Co-Authored-By: Claude Opus 4.8 <[email protected]>

47ed2c6b8b

• View at this point in history

2026-08-01 01:24:54 -04:00

Transaction Science

wai-rs: dispatch wai.image.jpegai to the deterministic decoder when its bundle ships ...

The envelope dispatched `wai.image.jpegai` to the old ORT float decode even after the
byte-exact decoder merged. Wire the deterministic path in, additively: when the model
directory carries the integer weight bundle (`weights/*.json` + `norm.json`, as
`tools/wai_jpegai_build_e2e_inputs.py` produces), decode through
`jpegai_decode::decode` — the reconstruction with a portable decoded-pixel hash. When
that bundle is absent, fall back to the ORT decode unchanged, so existing ONNX bundles
keep working.

`neural` already implies `neural_int`, so the deterministic module is reachable from
the dispatch. The file I/O lives at the dispatch boundary (`try_jpegai_exact`), keeping
`jpegai_decode` itself I/O-free; the weight/norm loaders reuse `IntNet::from_json`,
`Weights::from_json`, and a new `IcciNet::from_json` (+ `Conv`/`IcciPath::from_json`).

Tested (feature `neural`): with the bundle present the loader decodes the real
test256 codestream to a 256×256 RGB image; with it absent `try_jpegai_exact` returns
None so dispatch falls back. Builds clean under the ORT feature.

Co-Authored-By: Claude Opus 4.8 <[email protected]>

64fd36d98c

• View at this point in history

2026-08-01 01:24:54 -04:00

Transaction Science

wai: house-style pass on public copy — drop comparisons, vendor names, vanity counts ...

Bring the git-visible surfaces in line with the framing rules (self-referential;
no competitive comparisons, no vanity metrics, describe by capability):

- SPEC.md neural table: drop company prefixes (Meta/Descript-company/Kyutai → the
  capability's own model identifier), drop BD-rate "vs VTM/VVC" comparisons and the
  "current SOTA" superlative (describe each capability's regime instead), and reword
  the "vs the float X family that desyncs" lines as self-referential statements of
  why the integer contract exists ("a float decode desyncs on…").
- README status line: "passes 11 … tests" → "round-trips every registered capability
  through the full envelope" (a count that describes coverage, not a number that drifts).
- glossary: "canonical 342-entry dataset" → "canonical survey dataset".

Unchanged on purpose: "dispatches to SOTA standard codecs" and similar — that is the
ingest thesis (we route to the best existing codecs), self-referential, not a
comparison of our own work; and the source-column model IDs, which are factual
dependency pointers.

Co-Authored-By: Claude Opus 4.8 <[email protected]>

ea31d22854

• View at this point in history

2026-08-01 01:24:54 -04:00

Transaction Science

wai: register the byte-exact JPEG-AI decoder + restore the byte-equality matrix ...

Integration follow-through on the merged jpegai_decode: make the decoder visible
where the standard describes itself, and get it into the cross-arch matrix.

SPEC/docs:
- Register `wai.image.jpegai` in SPEC.md §5 (neural table) — a live, dispatchable
  capability was absent from the normative menu. Framed as decode-equivalence: one
  pinned reconstruction of an ISO/IEC 6048 codestream to RGB, reproducible on every
  machine (a portable decoded-pixel hash), entropy/dequant integer-exact, synthesis
  within 1 LSB of the reference.
- CHANGELOG entry for the byte-exact codestream→RGB decoder, distinguishing its
  decode-equivalence profile from the all-i64 hash-portable in-house codecs.
- Fix the stale CAP_IMAGE_JPEGAI comment ("decoder library not yet in this crate" —
  untrue post-merge) and drop its "~30% over conventional SOTA" comparison.

byte-exact-conformance:
- The crate did not compile on main: `rans.rs` was moved out of `neural/` to the
  wai-rs crate root, but the shim still #[path]-included the old `neural/rans.rs`;
  the moved file also references `crate::int_entropy`. Repoint rans to the crate
  root and shim `int_entropy` (self-contained). This restores the cross-arch
  byte-equality build.
- Add a golden that decodes the same ISO reference mANS stream through the MERGED
  decoder's entropy stage (crate-root `jpegai_mans`, flat-table/sliced-offset API)
  and asserts the identical symbols. The decoder `wai.image.jpegai` actually
  reconstructs with is now in the byte-equality matrix, not only the older
  `neural/` port. 15 conformance tests pass.

Co-Authored-By: Claude Opus 4.8 <[email protected]>

c4475c4ecb

• View at this point in history

2026-08-01 01:24:54 -04:00