PoC20 — trained SqueezeNet spatial compatibility

Date:2026-10-14. Owner authorized testing other real models and researching strong noise reduction/echo cancellation, with VAD integration deferred. The design and plan were committed before native work. Controller alone operated the device; REA workers reviewed source/evidence and researched audio candidates.

Result:20a spatial Conv/Relu,20b MaxPool and20c Fire/Concat qualified on NN. 20d’s exact classifier AVG13×13 kernel failed initialization during the first size-only export API call. It returned5 with clean health/cleanup, produced no NBG and performed no inference.20e and the conditional benchmark were skipped. No native retry, precision-policy relaxation or CPU neural fallback occurred. Actual budget use:4 compiler processes/7 public CompileToBinary calls/3 NPU Runs; planned maximum6/12/30.

The guide records numerical results, model provenance, bounded operator claims and the failed initializer. The machine-readable ledger binds model, source/ELF, NBG, run, reference and cleanup identities.

Evidence retained

• artifacts/npu/poc20/prepared/:32 frozen files and exclusive receipt; original trained FP32, explicit FP16 exports, four images’ inputs/references.
• independent-oracle/:eight separately retained NumPy outputs and actual ORT comparison; core-verification/: coefficient bits/source regeneration and unchanged older profile/source proof.
• packages/20a–20e:lowerings, exact compiler source/AArch64 ELF/build receipts;20a–20c have NBGs, executor and qualified evidence.20d has a retained clean negative;20e is host-built but never exported.
• artifacts/npu/runs/rea-poc20{a,b,c,d}-*-compile-20261014-a1 and rea-poc20{a,b,c}-*-execute-20261014-a1:raw command, result, before/after, numerical and cleanup evidence. classifier-negative-audit.md preserves the independent exact-source/backend attribution.
• packages/native-phase.json:35 operation events plus one final read-only continuity check. Literal terminal state ACTIVE; experiment closed and no native call remained. No automatic reset or final cosmetic advance.
• host-final-verification.json:three qualified pure loads, twelve copied tamper refusals and two unqualified strict/missing-NBG refusals, with ADB and ORT-session throw guards. No actual CPU neural session or device call.
• final-host-tests.txt:381 passed,5 subtests passed. Conditional benchmark RED/host-development logs retained, but unused source removed and no benchmark compiled or executed.

Health history

host-preparation-health-0001.json remains FAIL: the boot changed during host fixture/source preparation, before any SqueezeNet native call, and Wi-Fi diagnostics changed. Cause unknown. Temporary manual supplicant disconnect used prior owner authorization and did not lower wlan0 or stop services. The206.882-second quiet window and independently reviewed fresh current-boot seed admitted phase squeezenet-20261014-a1.

All native local and continuous phase checks passed. Every owned root was cleaned after successful execution or the clean rejection. The first final raw health comparison remains FAIL because its temporary-root CLC key differs from the seed’s key. final-native-health-0003.json binds the retained raw hashes and records the explicit identical SYSTEM_ABSENT projection and independent38-root absence check; all other hashes/state/diagnostics remain checked. Seed→final, seed→post-restoration and final→post-restoration comparisons pass with no new alerts. This does not reclassify the earlier boot/Wi-Fi failure.

wifi-restoration.json verifies the original saved network, COMPLETED, adminUP, assigned IP, ACK OK and no saved configuration. Private network details stay in ignored before/after files. No native calls followed restore.

The policy reviewer generated a deterministic local20a compile-manifest during host review, disclosed it and had the controller validate its bytes. It performed no ADB/vendor operation. Reports do not claim zero peer host artifact writes; source/evidence ownership remained separated.

Subsequent work

Isolate the AVG13×13 backend eligibility condition and prove any equivalent lowering before reconsidering full SqueezeNet. A clean negative is not an allowance for unselected retries. The next audio compatibility work requires new state/multiple-I/O/GRU/ConvTranspose/normalization controls; public ONNX availability does not qualify the Wave. See the primary-sourced speech-enhancement candidates. DPDFNet2_48khz_hr is the stronger full-band denoising candidate; LocalVQE v1.2 is a downloadable joint AEC/NS/dereverb candidate, DTLN-AEC128 an established control, and GTCRN the smallest first recurrent workload. AEC needs synchronized microphone/playback-reference pairs; received-speech denoising is a separate stream/state. No audio-model inference or integration was performed in PoC20. VAD integration remains deferred under C20.