wai: the int_codec audio trained on a generated signal, and every corpus states its provenance #184

Merged

dcharlot merged 10 commits from wai/audio-provenance into main 2026-10-07 21:26:28 -04:00

Conversation 0 Commits 10 Files changed 68 +1257 -249

dcharlot commented 2026-10-07 14:56:58 -04:00

Owner

Copy link

The learned integer audio codec's model and vectors are now trained on, and code, a generated signal. Earlier revisions used a third-party recording. Every corpus now states the provenance of its content.

• The audio.
  • tools/wai_int_audio_export.py generates its own signal from a seed: tones, chirps, noise bursts and formant sweeps. It is integer-only and takes no audio input.
  • The model, the vectors and every corpus that embeds them were regenerated through their own generators:
    • the int_audio vectors;
    • the int-payload WIA1 vectors and refusal fixtures;
    • the staged-delivery audio case;
    • the staged-measurement audio fixture and its goldens.
  • Training is deterministic: two runs give byte-identical weights. Rust's integer decode equals the Python reference.
• A guard.
  • tools/wai_audio_provenance_check.py correlates every committed audio vector against a reference set of recordings the CI host has, and fails above a stated threshold.
  • A host without the reference set reports the step as skipped, with the reason.
• Provenance everywhere.
  • Every corpus README carries a provenance line, and tests/vectors/README.md lists each vector directory's source.
  • tests/corpus_provenance.rs fails when either is missing.
  • The rule is in CONTRIBUTING.
  • Two vectors are recorded as not generated and are to be replaced in a following change: a real-footage video vector and a photograph vector.

Checked locally on macOS arm64:

• the link gate;
• every corpus through both implementations;
• the guard, falsified both ways;
• the public-text checks.

The forge CI runs on this pull request.

🤖 Generated with Claude Code

The learned integer audio codec's model and vectors are now trained on, and code, a generated signal. Earlier revisions used a third-party recording. Every corpus now states the provenance of its content. - **The audio.** - `tools/wai_int_audio_export.py` generates its own signal from a seed: tones, chirps, noise bursts and formant sweeps. It is integer-only and takes no audio input. - The model, the vectors and every corpus that embeds them were regenerated through their own generators: - the int_audio vectors; - the int-payload WIA1 vectors and refusal fixtures; - the staged-delivery audio case; - the staged-measurement audio fixture and its goldens. - Training is deterministic: two runs give byte-identical weights. Rust's integer decode equals the Python reference. - **A guard.** - `tools/wai_audio_provenance_check.py` correlates every committed audio vector against a reference set of recordings the CI host has, and fails above a stated threshold. - A host without the reference set reports the step as skipped, with the reason. - **Provenance everywhere.** - Every corpus README carries a provenance line, and `tests/vectors/README.md` lists each vector directory's source. - `tests/corpus_provenance.rs` fails when either is missing. - The rule is in CONTRIBUTING. - Two vectors are recorded as not generated and are to be replaced in a following change: a real-footage video vector and a photograph vector. **Checked locally on macOS arm64:** - the link gate; - every corpus through both implementations; - the guard, falsified both ways; - the public-text checks. The forge CI runs on this pull request. 🤖 Generated with [Claude Code](https://claude.com/claude-code)

dcharlot added 10 commits 2026-10-07 14:56:58 -04:00

wai: the int_codec export generates its own signal from a seed ... 7bb8547dee

tools/wai_int_audio_export.py no longer reads an audio file. It writes the
training signal and the test signal (the clip a vector codes) itself, from a
seed: 16 kHz mono s16, segments of tones, chirps, noise bursts and
speech-like formant sweeps, sometimes two overlaid, with short gaps. The
generator is integer arithmetic only (a SplitMix64 counter stream, a 32-bit
phase accumulator, a Bhaskara I sine table, an exact 2^(k/64) table), so a
seed gives the same samples on every machine. The codec trains on stream 0
of the seed and codes stream 1.

Training is seeded from the same seed with deterministic algorithms and a
fixed thread count. --signal-only writes one generated signal with the
Python standard library alone. The model, payload and capability are
unchanged; the vectors are regenerated in the commits that follow.

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

wai: a check that no committed audio vector is a coding of a recording ... a16ea1bd5a

tools/wai_audio_provenance_check.py (Python standard library) reads every
committed audio vector to PCM — each audio base's stage 0 in the staged
corpus, each staged-measurement source over one, each spatial-audio source,
every audio file under wai/ — and computes its normalised cross-correlation
at every lag against every recording in a reference set the host has (the
host's system sound directory, --ref, WAI_AUDIO_REFERENCE_DIRS). Above the
threshold (0.5) it fails. A learned-codec vector is covered through its
decode, which must be a staged case's stage 0. With no reference set, or no
converter to read it, the check prints why and is skipped.

On this commit it fails: the int_codec stage 0 and the measurement source
built on it match a reference recording (0.984 and 0.963). The vectors are
replaced next.

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

wai_int_payload_vectors: the WIA1 table fixtures read the table's place from the header ... 91e76cc275

wia1_cdf_malformed and wia1_order_table_before_length change the last entry
of the payload's first CDF table, and the second cuts the payload inside the
second table. Both used fixed offsets for the old vector's shape (16
channels, tables of 14 entries). They now read n_ch (offset 12) and cdf_len
(offset 28) from the payload itself: the first table starts at 32 + 8·n_ch,
its last entry is at (cdf_len − 1)·4 from there.

On the committed vector the offsets are the same, so the corpus is
unchanged: this binary verifies the corpus of the previous commit byte for
byte. On a payload whose tables have another length the fixtures keep their
meaning, which the regenerated audio vector needs (its tables have 15
entries).

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

wai: int_codec model and vectors regenerated from the generated signal ... 9d31aad609

The wai.audio.int_codec vector (wai-rs/tests/vectors/int_audio/) is now an
export of tools/wai_int_audio_export.py on its own generated signal:

    python tools/wai_int_audio_export.py <out> --seed 1 --channels 16 \
        --samples 8192 --steps 2500

Two runs with seed 1 gave byte-identical audio_model.bin,
audio_bitstream.json, audio.json and decoded PCM. Every file that depends
on the vector was regenerated through its own generator, from it:

- int_audio/sha256.json — WAI_WRITE_SHA256_SIDECARS=1 cargo test
  --test int_kernels_vectors int_audio_clip (the Rust decode equals the
  Python integer reference's PCM first);
- int_audio/audio_bitstream.wia1, int-payload-conformance/vectors.json and
  the thirteen wia1_* refusal fixtures — wai_int_payload_vectors gen;
- staged-conformance/cases/audio_int_codec_1l/ (base.wai, L1.wai,
  stage0.bin, case.json) and index.json — wai_staged_vectors gen;
- staged-measure-conformance/fixtures/audio/ (source, layers, references,
  expect.json, title.json), goldens.json and sim/results.json —
  wai_stage_measure gen-fixtures, then generate-goldens. quality.json is
  regenerated unchanged (it holds no audio).

The format, the payload layout and the capability are unchanged; the
model's shape is the same (16 channels, 512 latent samples, 8192 output
samples at 16 kHz). Verified: the six affected Rust corpus verifiers
(int-payload, staged, staged-measure, http, claims, pin), the Python
verifiers int_payload_verify.py, staged_verify.py, measure_verify.py,
sim_verify.py, http_verify.py and pin_verify.py, and the integer-kernel,
int_audio, int_payload and stage_measure tests. The provenance check now
passes: the stage 0 and the measurement source correlate at most 0.22
with any reference recording.

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

ci: no committed audio vector is a coding of a reference recording ... a713028f86

The cargo-class job's unix legs run tools/wai_audio_provenance_check.py.
On the macOS leg the reference set is the host's system sound directory and
the check runs; the linux legs have none and print that the check is
skipped. A change to the tool fires the workflow.

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

Merge remote-tracking branch 'origin/main' into wai/audio-provenance 3aa59b86f1

wai: the int_codec vector's README, and the statements about what it was trained on ... a8bf1c560c

wai-rs/tests/vectors/int_audio/README.md states the vector's provenance (a
signal the export tool generates; no audio input), the seed (1), the
command, the tool versions, and that two runs gave byte-identical files.

research/decode-throughput.md keeps its text (a dated record) and gains a
dated correction: the int_audio vector it timed was trained on, and coded,
a third-party recording, and has been replaced by one trained and coded on a
generated signal of the same shape. wai_decode_bench's doc comment now lists
int_audio among the vectors trained on generated content.

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

wai: every corpus states the provenance of its content ... 3fac029a71

Each *-conformance/README.md now has a line that begins **Provenance.**:
the generator that writes the corpus and its input, and for an input not
generated here its source and licence. OER/2's corpus is settled and left
as it is. byte-exact-conformance and quantum-ops-conformance get a README.
wai-rs/tests/vectors/README.md gives a row per vector directory. Where the
staged-delivery, staged-measurement and telemetry READMEs called their
content "synthetic", that sentence is now their provenance line.

Two vectors are recorded as not generated, from the provenance audit:
int_video's video_real (a pan over a third-party picture) and
neural-conformance's photo (a third-party picture, with a decoder trained on
unrecorded images). Their lines say so and that they are to be replaced;
the photo bullet no longer names the picture's source product.

tests/corpus_provenance.rs requires the line in every corpus and a row for
every vector directory; CI runs it with the carriage tests. CONTRIBUTING
states the rule: corpus content is generated here, or recorded with a
licence that permits redistribution, and a word like "synthetic" is not a
provenance line.

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

wai: CHANGELOG — the int_codec model and vectors, trained and coded on a generated signal ... de0165bb34

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

Merge remote-tracking branch 'origin/main' into wai/audio-provenance 3a3c0a49d9

dcharlot merged commit 1cf132ecbe into main 2026-10-07 21:26:28 -04:00

dcharlot referenced this pull request from a commit 2026-10-07 21:26:30 -04:00

Merge pull request 'wai: the int_codec audio trained on a generated signal, and every corpus states its provenance' (#184) from wai/audio-provenance into main

Sign in to join this conversation.

Reviewers

No reviewers

Labels

Clear labels

No items

No labels

Milestone

Clear milestone

No items

No milestone

Projects

Clear projects

No items

No project

Assignees

Clear assignees

No assignees

1 participant

Notifications

Subscribe

Due date

The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set

Reference

Transaction-Science/open-standards!184

No description provided.