Wave NPU Erf-Mask Completion Audit

Audit date: 2026-07-29

Scope: design criteria 1–13 from docs/superpowers/specs/2026-07-29-wave-npu-learned-erf-noise-suppression-design.md

Verdict

The research objective is proven: the Insta360 Wave NPU numerically executed the learned Erf partition, its returned tensors were used in a causal noise-suppression pipeline, and the resulting complete waveform passed the quality and CPU-agreement gates.

The production milestone is incomplete. The exact model has not been ported behind the Rust NpuInfer service, and the fixed-runner soak is not a substitute for end-to-end service percentile, memory, reset, CPU-load, and thermal qualification.

research NPU noise-cancellation PoC: PROVEN
production Rust NPU backend:         INCOMPLETE
arbitrary ONNX or Acuity support:    NOT CLAIMED

Requirement-to-Evidence Audit

#	Requirement	Authoritative evidence and SHA-256	Observed result	Classification
1	Deterministic learned 962-parameter package	`artifacts/npu/runs/poc13a-erf-mask-host-20260729-model-v4-selector/model-manifest.json` — `bd7ffc58d253855e3ddc546def711f91812da8cb9ecd88dc658b56aada40ce5e`	481 slopes plus 481 biases; deterministic weights and ONNX are hash-bound	proven
2	Held-out ablation and quality gates	`artifacts/npu/runs/poc13a-erf-mask-host-20260729-evaluation-v5-selector-final/evaluation-summary.json` — `32d7b224253e2a05f230555db7cfdf4bb6094c81d832d5793851a582a376ee2d`; host qualification — `ad286a54ca1c3311cb1a44818c777de2dd62dca1b8d155b715e076f03710693d`	1,230 valid items; synthetic SI-SDR +3.2598 dB, natural +5.2338 dB, STOI +0.01355, noise attenuation 10.5838 dB, held-out clean RMS delta 0.04638 dB	proven
3	Exact U8 `[481]` Erf NBG compiles twice identically	A1 run — `aa9a6c860c273f0232fdabe57238b5c1911bb7cf3044fea5bad600e532725647`; A2 run — `de57fcf8a8f88211255f979267c7802808913562b1e04234f9b3e6a4b7d8ea66`; both NBGs — `96447d6afd63a7c349bbd80f0af99150c89ef24ed7f1ef66ab285075362e6c94`	Two independent compile-only processes produced byte-identical 5,840-byte VPMN v2 files for `0x1000000e`	proven
4	Strict parser, CRC, metadata, signature, and corruption checks	`docs/npu/format/erf-mask-nbg-profile.json` — `9e5300363c5eca8bd9dae6db820128d4e9534f2aa6007a69fe10bf7b9bc9cab7`	CRC `0xf14f321a`, exact U8 tensor metadata, lossless ownership, zero conflicts, structural signature `236bd56d…`; negative tests pass	proven
5	Six bounded NPU vectors match CPU oracle	`artifacts/npu/runs/poc13d-erf-mask-six-vector-20260729/run.json` — `f8ffbca89fb96275e1c404ea780c1ccb62de3deda07820ec7308ccf6aaab7fac`	2,886/2,886 bytes exact; max error 0; learned output has 63 unique values; canaries pass	proven
6	Strict numerical calls reached NPU without CPU fallback	Same PoC 13D run and fixed-runner policy; runner binary SHA-256 `99d51397b8c6b21c360755b6a66cc62eba3897287fb12a215468bd8f7e2f62d3`	One TIM-VX graph compiled, six NN SDK executions ran on the identified Wave target, and the runner contains no CPU Erf implementation or fallback path	proven
7	Complete waveform reconstructed from NPU-returned masks	Stream run — `6079c34599386f2024ac7787deacf01413231b5bf312f75b030292f2cb87bda2`; waveform evidence — `6589c517483fbf120e70176c46267c867f96dc2d0286d57521920dab2d93d09b`	1,201/1,201 NPU frames, 577,681/577,681 bytes exact, 576,000 samples reconstructed, zero fallback	proven
8	Waveform passes quality and CPU/NPU agreement gates	Same waveform evidence; rendered WAV SHA-256 `30e015ed85faacd5198c70a3021fd59557103945bd19c24b496a58f4ee21286c`	Noise attenuation 9.8283 dB, SI-SDR +4.8443 dB, STOI +0.01083, clean transition 1.1822 dB within its 1.5 dB fixture gate, silence −85.6481 dBFS, CPU/NPU samples exact	proven
9	Production implementation runs behind Rust `NpuInfer`	No qualifying artifact	Python/lab integration and fixed native device runners exist; no Rust service backend was added	incomplete
10	Production strict, hybrid, CPU, and audit modes	Python fallback source SHA-256 `001295b27aea1b5d7a4547b1f78d7fcf49257aa370e7ba741a5da2d6f05240e4`; exact partition policy SHA-256 `f5c7d9a76e44a20e10d0af0f7e7cfba76c0de99a8cac616eb22b471f079de44d`	Strict fixed-device execution and Python CPU/hybrid behavior are proven; the four production service modes are not wired	incomplete
11	Same-frame fallback and no automatic retry	`waved/tools/onnx-hybrid/tests/test_erf_mask_device_stream.py` — `3b0d9dd9dfb3d60096c26e53bc3e365a1783ab8de5b56420c08548b3c9e080ff`; waveform evidence above	Injected first NPU failure recomputes that frame exactly on CPU, latches fallback, and makes only one NPU attempt	proven at lab boundary
12	10,000 frames meet all latency, ordering, memory, health, and cleanup gates	`artifacts/npu/runs/poc13f-erf-mask-soak-10000-20260729/run.json` — `e4f3fc3d0de3c833dadf912563585c3737d7dc30c7338f2e26d9936077ee6f8d`	10,000 ordered executions and 4,810,000 bytes covered; 10/10 oracle checkpoints; mean 84.8758 µs, max 1,649 µs; canaries and health pass. No p50/p95/p99, service RSS/CPU, reset, or thermal evidence	incomplete as written; numerical/health subset proven
13	Manifests, notebooks, hashes, and reproduction commands agree	This audit, `README.md`, four tracked JSON manifests, six PoC 13 notebooks, and the hashes rechecked below	Required research evidence is documented; explicit production non-claims are retained	proven

Reproduction and Verification

Host model and hybrid tests:

cd waved/tools/onnx-hybrid
uv sync --frozen
uv run pytest -q
uv run ruff check .

NBG parsers, policies, and device-harness tests:

uv --project waved/tools/onnx-hybrid run pytest -q waved/scripts/npu/tests

Static policy checks:

waved/tools/npu-compiler/verify-device-erf-mask-policy.sh \
  artifacts/npu/build/poc13b/wave-npu-compile-erf-mask-v1 \
  waved/tools/npu-compiler/tim_vx_compile_erf_mask.cc

waved/tools/npu-compiler/verify-device-erf-mask-qualification-policy.sh \
  artifacts/npu/build/poc13d/wave-npu-run-erf-mask-qualification-v1 \
  waved/tools/npu-compiler/wave_npu_run_erf_mask_qualification.cc

waved/tools/npu-compiler/verify-device-erf-mask-stream-policy.sh \
  artifacts/npu/build/poc13e/wave-npu-run-erf-mask-stream-v1 \
  waved/tools/npu-compiler/wave_npu_run_erf_mask_stream.cc

waved/tools/npu-compiler/verify-device-erf-mask-soak-policy.sh \
  artifacts/npu/build/poc13f/wave-npu-soak-erf-mask-v1 \
  waved/tools/npu-compiler/wave_npu_soak_erf_mask.cc

Evidence-hash verification:

shasum -a 256 \
  artifacts/npu/runs/poc13a-erf-mask-host-20260729-model-v4-selector/model-manifest.json \
  artifacts/npu/runs/poc13a-erf-mask-host-20260729-evaluation-v5-selector-final/evaluation-summary.json \
  artifacts/npu/runs/poc13a-erf-mask-host-20260729-qualified-v4-selector/host-qualification.json \
  docs/npu/format/erf-mask-nbg-profile.json \
  artifacts/npu/runs/poc13d-erf-mask-six-vector-20260729/run.json \
  artifacts/npu/runs/poc13e-erf-mask-fixture-stream-20260729/run.json \
  artifacts/npu/runs/poc13e-erf-mask-waveform-v2-20260729/waveform-evidence.json \
  artifacts/npu/runs/poc13f-erf-mask-soak-10000-20260729/run.json

The device stage scripts refuse pre-existing remote roots and intentionally have no retry or recovery mode. Existing /tmp/wave-npu-poc13* evidence must not be deleted merely to replay a run.

Exact Claim Boundary

Supported:

custom fixed TIM-VX graph generation for this quantized Erf contract;
VPMN v2 parsing, hash binding, loading, and numerical execution;
a real enhanced waveform materially influenced by Wave NPU output;
partition-first hybrid execution with deterministic CPU fallback;
sequential 10,000-frame numerical stability for the fixed partition.

Not supported:

arbitrary ONNX import or full-model NPU residency;
a complete Acuity compiler/profile;
general Conv, MatMul, FFT, recurrent, or attention execution;
a production-ready Rust/audio-service integration;
a production real-time or thermal qualification claim.