657daba14a
WAI named two of the DCVC neural video codecs and none of the rest, nothing for the neural-network-based video coding exploration of ITU-T and ISO/IEC, and no lossless frame format for tensors. SPEC §5 registers wai.neural.dcvc, .dcvc_tcm, .dcvc_hem, .dcvc_dc and .dcvc_uf as candidates (neural-float, media class video, patent-pool provisionally: the code is under MIT, which grants no patent licence). Each published decoder takes its latent's entropy parameters from a temporal context as well as the hyperprior, most from a spatial prior of two or four passes too, four decode a motion latent, and DCVC-UF codes a chunk of frames as one latent; none fits the inter-frame route of neural/video_dcvc.rs, whose latent comes from the hyperprior alone, and none pins a bitstream. A new SPEC note says so, and that dcvc_rt and dcvc_fm decode through that route with a sink-supplied temporal decoder. wai.video.nnvc is a candidate too: no standard is published and the reference software's bitstream follows its release. Candidates are refused by dispatch and never picked (SPEC §7.1), which the existing candidate tests now cover for all six. wai.tensor.weft registers one frame of the WeftCodec format 1.0 (WEFTCODEC-FORMAT v1.0.0, WEFT-AI-CODEC v1.3.0, weftcodec 1.2.0) in mode RAW, TENSOR or FLOAT_TENSOR, naming no dependency: lossless, its plaintext bound by a BLAKE3 content hash, so standard-defined; royalty-free (text CC BY 4.0, reference implementation Apache-2.0). Any other mode is UnsupportedMode, as the format requires of a decoder that does not implement a mode. The reference is codecs::weft (feature weft): a decoder that checks a frame in the format's order and refuses it with the format's own class, and writers that give the format's own frames byte for byte. The corpus, tensor-conformance/, holds 77 of the format's frozen vectors with the outcome this capability gives each, and 70 generated frames, five that decode and the rest with one planted defect each, every class and every limit reached. weft_verify.py, its second implementation, agrees on all 147, refuses all 143 changed frames, and requires each of its 77 checks to refuse some vector and to be needed: 72 alone, five (always backed by another check with the same class) together with that check. staged_verify.py's video class names the DCVC candidates as the reference's does. The registry export, REGISTRY.md and wai.h are regenerated. Thirteen faults planted, each failing its tests: DCVC-HEM not a candidate, NNVC royalty-free, DCVC-UF outside the video class, ZSTD admitted, no content hash, no header CRC, a 25-bit code admitted, no exponent range, no table_len check, weft as neural-float, SPEC missing a class (both implementations), and the second verifier without its exponent range. Co-Authored-By: Claude Opus 5.5 <[email protected]>
3.3 KiB
wai.tensor.weft conformance corpus
The vectors of wai.tensor.weft (SPEC §5, WFT1 tensor frames): frames of the WeftCodec format
1.0 (WEFTCODEC-FORMAT v1.0.0; WEFT-AI-CODEC v1.3.0 for its tensor modes; as published with
weftcodec 1.2.0), each with the outcome a sink that advertises the capability gives it: its
plaintext's length and SHA-256, or the format's error class it is refused with, under the limits
the vector names.
Provenance. Two sources. The frames marked third-party in vectors.json are the format's
own frozen conformance vectors (conformance/v1/codec/ of its reference implementation, at
weftcodec 1.2.0, commit d95fa78c805e, licensed Apache-2.0), copied unchanged by
tools/weft_import.py: every one but the vectors a later vector supersedes and the one-mebibyte
RAW vector. A frame of a mode the registration admits (RAW, TENSOR, FLOAT_TENSOR) keeps the
format's expectation; any other is UnsupportedMode here. The frames marked generated are
written by wai-rs/src/codecs/weft/corpus.rs from fixed integer sequences: five that decode
(an i8 tensor, fp32 weights, bf16 NaNs, infinities, zeros and subnormals, an fp8 plane of one
exponent, a RAW frame in three chunks) and frames with one planted defect each, every other field
written to agree with it, so that only the rule the defect breaks refuses the frame (several are
built so that a decoder without that rule would hand over a plaintext). licences.json records
each vector's licence and source.
| file | what it holds |
|---|---|
vectors.json |
every vector: its file, the SHA-256 of its frame, its limits, its outcome; a third-party vector also the format's own expectation, a generated one what it plants |
licences.json |
each vector's licence, content and source (scripts/check-wai-corpus-licences.py) |
vectors/<name>/frame.wft |
the frame |
weft_verify.py |
the second implementation |
Regenerate and check (from wai/wai-rs/):
cargo run --no-default-features --features weft --bin wai_weft_vectors -- gen ../tensor-conformance
cargo run --no-default-features --features weft --bin wai_weft_vectors -- verify ../tensor-conformance
weft_verify.py is written from the format's text and SPEC §5 and shares no code with wai-rs:
its own frame reader, CRC-32C, canonical structured-value reader, canonical Huffman decoder and
plane join, and the BLAKE3 of claim-conformance/_blake3.py. It requires every frame to have the
digest vectors.json states and every vector its outcome; the classes and modes vectors.json
lists to be SPEC's; every frame it decodes to be refused once any one of several of its bytes
is changed; and each of its own checks to refuse some vector and to be needed, by dropping each
in turn and requiring some vector to change its outcome. A check that another always backs with
the same class, for every frame of an admitted mode (a domain above 1, which the admitted modes'
domain rule also refuses; a dependency, which no admitted mode names; a RAW frame's chunk
extents and its chunk sums, each of which implies the other), is kept so that a frame with
several defects is refused in the format's order, is listed with the check that backs it, and
is required to change some vector's outcome when dropped together with that check. Standard
library only:
python3 -I tensor-conformance/weft_verify.py tensor-conformance