rUv
aac4155584
* feat(bfi): add validated capture and bounded research tooling * fix(deps): patch Wasmtime for RUSTSEC-2026-0316
5.9 KiB
ADR-365: Public BFI dataset and decoder contract
- Status: proposed
- Date: 2026-09-29
- Deciders: RuView maintainers; acceptance pending
- Tags: bfi, datasets, decoding, provenance, benchmarks
Context
Public activity recordings offer a separate route to BFI model development while local capture continuity is insufficient. They introduce different frame formats, collection assumptions and licenses. Public HAR activity labels do not reproduce KIT's person-identification experiment.
MEASURED: the pinned CSI-BFI-HAR train/validation download passed exact size/hash checks for 120 captures plus its license card. A separate body decoder matched 13,104 quantized angles across 14 sample bodies against tshark 4.2.2 in two independent executions. Whole-packet admission and activity training remain incomplete. See results and the decoder audit and reproducer.
Decision
Pin provenance and partitions
Use the author dataset
at revision c3013b3edc5a563a9cf6759f5d464a9d9b6f456a. The
manifest
pins URLs, lengths, SHA-256, labels and split membership. Preserve the pinned
dataset card's GPL-3.0 declaration and attribution separately from code licenses.
| Partition | Author directory | Domain |
|---|---|---|
| Training | HAR-1/BFI/M1 | Day 1, kitchen |
| Validation | HAR-3/BFI/M1 | Day 3, classroom |
| Held out | HAR-5/BFI/M1 and M2 | Day 5, living room; two views |
Keep all views of one event in the same partition. Day and room are confounded; M1/M2 labels do not establish heterogeneous-chipset transfer. The selected filenames cover A–T activities and P1–P3 participants.
Default downloads exclude held-out views. The bounded downloader verifies exact bytes before atomic completion, rehashes resumed files, retains failed partials and enforces transfer deadlines and byte limits. Data stays outside Git; downloaded code is never executed by the downloader.
Keep body decoding separate from packet admission
public_bfi_decode.py accepts only the audited VHT MU 3x1, 80 MHz, Ng=1, codebook-1 complete action body: 1,003 bytes containing one SNR byte, 936 angle bytes and 61 opaque MU-exclusive bytes. It exposes quantized/radian angles and tone ordinals. Frequency indices, exclusive-field interpretation and validated matrix reconstruction remain separate acceptance work.
Twelve sample frames have nonzero fragment-number bits. Successful tshark decoding and valid CRCs do not resolve their semantics. Retain live MAC guards; quarantine these frames until an explicit, independently reviewed dataset policy can justify admission. Require lossless pcapng conversion evidence for integer timestamps, packet lengths, link types and all original packet bytes. Do not silently strip headers or rewrite fragment fields to gain admission.
The natural circular representation needs 1,404 features, exceeding the current trainer's 1,024-feature limit. Approve a bounded representation or revised limit with allocation tests before integration; do not truncate tones implicitly.
Admit recording quality before scoring models
Inventory every manifest file, including tiny or unusable traces. Freeze a five-second observed-interval rule: at least 20 valid unique reports, with no inter-report gap above one second and no interpolation. Require at least two independent recordings per activity in each partition/view group that supply an eligible interval. Missing A–T coverage fails benchmark admission. Before implementing the census, also freeze interval anchoring, stride and leading/trailing gap rules: an inter-report gap bound alone does not bound silence at interval edges. These details remain required acceptance work.
Public files lack trusted recorder start/end manifests. Observed packet spans cannot establish unobserved leading/trailing coverage. Report rejection reasons, eligible recording/time coverage and accuracy together. Freeze event/file groups before generating windows; inspect held-out quality without model selection against held-out outcomes.
Alternatives
Reusing the HE/SU parser without format admission, weakening live fragment guards, or silently dropping sparse classes would change the task while hiding the change. Keep this adapter explicit and fail incomplete benchmark coverage.
Consequences
Positive
Data provenance, protocol support and score coverage remain independently auditable. New public formats cannot silently weaken live capture validation.
Negative
Sparse files or unresolved framing may prevent a benchmark despite a successful download. Format and representation work must precede model optimization.
Neutral
This is offline research/reference tooling. Production integration must reuse the Rust protocols in ADR-291 and scorecards in ADR-317; this ADR does not establish a second production trainer or an integrated benchmark score.
Validation and links
Use test_download_public_bfi.py and test_public_bfi_decode.py under
tools/bfi_learning, then the full packet oracle and quality census specified
in the decode audit and
benchmark protocol.