2026-10-13 — PoC19: MarbleNet convolution compatibility

Completed after owner reconnection on2026-10-14. The complete trained graph qualified on NN/TP, all800 frozen frame decisions match FP32, and the paired prepared-feature benchmark measured10.399× faster boundary execution. Wi-Fi and final device state are verified restored. The failed2026-10-13 and first2026-10-14 intervals below remain failed. See the resumption section.

Scope and identity

The owner authorized continuing the REA investigation, a compatible trained model rewrite and staged native testing. The controller owns every device operation. Source reconstruction, native static analysis and independent review have separate scopes. The existing feature checkout and unrelated changes are preserved. No production service deployment is implied.

This continues PoC18, preserving its rejected K29/rate2 control and all earlier receipts. Checkpoint 84bda37e925ac6fd740c2ced55642cb79f94f81348e1fa0db992ca50d4b09706, TIM-VX 702715d714277376d7ff466b7b72833ff57564c1, runtime c984e063e9ab4fe4d7e7ac85e783d88a330f1560268924ba4f449f5b258478dc and the pinned offline cross image are unchanged. The current boot identity was already different from PoC18 before this experiment began; no reboot or recovery was performed by this experiment.

Isolation results

Control	Trained-layer change	NBG bytes	Actual native result
19a	centered13 original taps, rate2, pad12	24,608	NN compile and one Run PASS; max error0.000244140625
19b	all29 original taps, rate1, pad14	26,656	NN compile and one Run PASS; max error0.000244140625
19c	exact Pad28/28 then original K29/rate2 VALID	32,168	TP Pad + NN Conv; one Run PASS; max error0.000244140625
19d	original first K11/stride2 block, flat-height native2×2	28,960	NN compile and one Run PASS; max error0.001953125
19e	original K15/K17 residual blocks	79,376	local NN compile PASS; surrounding health interval FAIL; no inference
19f	full exact Pad graph	none	not launched

19a/b reuse the exact frozen18C activation and downstream pointwise/bias/Relu. They are diagnostic graphs, not equivalents of the complete trained layer. The14-file preparation receipt is 692d3e2499fa447651f7296c660cf3f0c8cacd4f6bb3a677e927e31f8030f204. Run names are rea-poc19{a,b}-isolation-{compile,execute}-20261013-a1. All outputs are finite, with zero failures under the unchanged 0.02+0.02*abs(reference) element policy. CPU ORT, independent NumPy and actual native outputs are retained separately. Health/alerts were empty and cleanup was CLEAN. Neither isolated width nor isolated dilation explains the prior combined K29/rate2 rejection.

REA evidence and limits

An owned REA6.4.0/Ghidra12.1.4 session inspected the exact firmware OpenVX library, SHA f9cd917d4232316ab53a443ec93cd1217893f4de96fc113811b800b9ae8ca0bb. It was closed, its process/temp directory removed, and the target hash remained unchanged. The retained113-record evidence bundle is3,473,580 bytes, SHA b3321361bee1281a3bd9b58ab9a71f9165eaa07a94f9d70f05e5aa64bc1afa75. Addresses below are REA/Ghidra default-space addresses, rebased by0x100000. The decisive width load maps to measured file offset0x25d2c4, and the status−1 instruction maps to0x27c648; raw instruction evidence is retained.

• vxConvolutionLayer at0x24b200 selects kernel0x1000/table0x70a0f0 and initializer0x35cfa0. Its trivial validator0x362500 is not that initializer.
• Generic depthwise admission depends on representable signed padding, batch/format/weight flags, multiplier, runtime-configured kernel bounds, dilation support, and either unit strides or a feature-gated2×2 stride. The width comparison at0x35d2c4 uses context+0x2e78e8. These live bounds come from gcoVX_GetNNConfig; their numeric values remain unknown.
• The specialized depthwise fallback has unconditional false NN/TP predicates; EVIS admission is conditional. Once an admitted backend’s initializer fails, this dispatch returns that failure rather than trying another backend.
• The FP16 shader getter includes width/dilation branches. A null shader at 0x37b2d4 can reach status−1 at0x37c648; reshape failures can also return−1. The historical18C log cannot identify which branch actually fired.
• Static public TIM-VX Pad reaches native NN/TP routes. One-input static Slice can reach StridedSlice/view/copy routes; two-input Slice has different routes. This source eligibility does not qualify any new fixed graph.

Missing depthwise shader resources were observed in the staged SDK and matching searched firmware/SDK trees. Their absence and CLC quarantine do not establish the cause of18C. No arbitrary15-tap limit is inferred.

Selected exact rewrite

The first candidate explicitly pads [1,64,200] by positive-zero28 samples on each temporal side, producing [1,64,256]. The unchanged original K29/rate2 depthwise filter then uses VALID pads0 and produces200 frames; pointwise/bias/Relu remain unchanged. Effective extent57 gives256−57+1=200. Every original initializer bit is retained.19c directly established native TP padding plus NN convolution and passed finite numerical qualification.

A separate fp16-marblenet-pad-v1 profile binds structural constants, canonical Pad geometry and a flat-height depthwise stride policy. For native height1/kernel-height1/height-pad0 only, temporal stride2 uses native2×2 strides with the same height1 output.19d qualified that policy independently. Full-model admission reverses the deterministic rewrite to the exact original FP16 model hash; shape alone is insufficient.

Health and temporary Wi-Fi change

Three real atbm Wi-Fi WARN/tx-flush incidents occurred before any new model/NPU call, and the original failed snapshots are retained. A separate false-positive matching hang inside changed was fixed with regressions; actual warning/timeout detection remains intact. The owner explicitly authorized temporarily disabling Wi-Fi and restoring it afterward.

The controller captured original state privately, requested supplicant disconnect at2026-10-13T13:12:04.475Z and lowered wlan0. The running waved poller raised the interface again, while manual supplicant DISCONNECTED persisted. Credentials/configuration, stock services/watchdog, NPU settings, driver, firmware and system libraries were unchanged. A separately reviewed quiet71-second interval established the new eligible measurement phase. The prior failing intervals are not reclassified as healthy. Final Wi-Fi restoration and closure evidence remain required.

USB incident and2026-10-13 stopping point

The literal WARNING: Drop the transfer event appeared at uptime4992.636165, 37.14 seconds after19d CLEAN cleanup at4955.50. It was already in19e’s stage-before at5008.78 and preflight at5016.07.19e local before/after PASS therefore means no further change after its own baseline; it does not clear the genuine failed interval from the previous accepted state. Snapshot0015 has exact diagnostic overlap, unchanged static state and this new alert. Native launches stopped when the encompassing interval check found it.

USB_STATE=DISCONNECTED follows the warning by25µs, with gadget reconfiguration nearby. The public Khadas Amlogic5.4 CRG UDC source contains the same text when a transfer event refers to a dequeued USB request. This supports a USB-associated explanation, but it is not an exact binary callsite proof for this Wave kernel and does not reclassify the failed interval.

The controller attempted the already-authorized Wi-Fi restoration, but ADB reported the selected device absent before the first link command succeeded. No reassociation command was issued and restoration is unverified. Host ADB inventories continue to show no device. The owner was asked to reconnect with debugging enabled and clarify whether USB was changed. No driver reset, USB configuration mutation, daemon restart or new native model call followed.

Independent review3 identifies a real workflow gap: continuity from the last accepted cleanup was not checked before19e staging, allowing the new warning into its local baseline. Commit0843080 adds a persistent host phase guard for the new Pad profile. It checks the accepted anchor before staging mutations, after staging and on immediate native preflight; warnings/lost overlap/state changes latch HALTED. Only verified CLEAN cleanup advances the anchor. A narrowly validated clean compiler negative advances then rethrows its negative result; faults/numerical/artifact failures remain halted. Local NoRetry and missing-NBG requests refuse before phase activation. No production phase or seed has been created. This supports the sole controller workflow; its owner UUID is not an interprocess lock or compare-and-swap mechanism.

Status at the2026-10-13 stop

The bounded plan permits at most8 compiler processes,16 size/export API calls and46 NPU Runs, without retrying a control. Actual use is5 compiler processes,10 size/export API calls and4 NPU Runs. No control was retried. Compile deadlines15s native/20s host and execution8s/12s cover teardown. Faults, timeouts, numerical failures, ambiguous exits or health changes stop further native work; a clean compiler negative can guide another independent, reviewed candidate within the remaining budget.

The padded block and first trained stride2 block passed native qualification. Remaining K15/K17 inference, complete graph/four clips and same-device benchmark were not launched after the incident. Every completed caller’s local health/cleanup checks passed; the encompassing interval failed and final target state cannot be verified while ADB is absent. No full-model native success or speedup is claimed. The sanitized ledger retains these distinctions.

The frozen55-file Pad preparation is 92019faefcd00d93faee1d28b5d2cf9c8b1ba63d8c13bc71ac00048ac79df76d. Full candidate SHA 4929d1c702ed471bc6a9c50fae46bd23561892457405b4fd02fffb372ac10186 retains all34 folded initializer tensors,90,418 elements and180,836 FP16 bytes. Four-clip folded FP32 and FP16 logits are byte-exact to the originals; all800 decisions match the original unfused FP32 reference, probability error0.0007275641. Independent FP16 comparisons pass with max error0.00390625.

Source commits are core589fe0b, nativebce86be/16d0a4b and admissionb427b78. Root harness427 tests passed with2 optional skips; relevant non-ERF host tests430 passed. The broad host collection retained a negative for9 unrelated ERF modules needing SciPy/pystoi absent from the isolated Torch environment. Pure-host full admission accepted the exact rewrite and refused an altered weight even after resealing its digest, with zero CPU sessions/transport. The real unqualified full frontend refused before features/session/ADB. Actual qualified19c/d and valid copied packages load without CPU/transport; 26 altered-artifact cases refuse under CPU-session/ADB guards. The original model, source, output and evidence files were not modified by these tests. The guard’s full root harness passed470 tests, with2 optional skips and329 subtests. The last uncompiled-qualification ordering correction has a separate RED→GREEN refusal regression, with zero phase/oracle/transport/attempt marker. Final host verification rechecks all123 prepared files and14 packages, including10 qualified ones, with CPU-session/ADB guards. The sole-controller guard passed independent review; this does not authorize live continuation or replace the outstanding Wi-Fi restoration and final device checks.

2026-10-14 resumption and complete native qualification

The owner explicitly reconnected the unit. Boot identity is now eff903fb-00b5-45ab-bcad-30f36195cef5; the pinned serial, kernel, NPU settings, system libraries and model/build/preparation hashes match. No agent reboot, driver reset, service restart or USB configuration change was used. Wi-Fi was already COMPLETED on the original saved network at reconnection.

The first explicitly reviewed new phase, a1, then HALTED before staging or any vendor call: a fresh atbm wsm_write_mib WARNING at uptime363.952494 and lost200-line diagnostic overlap were detected. Its exact HALTED document, SHA ace9b39d0a235121a47a33b38a87dc51f4375b6a139016a76209d36c67a88dbc, and seed remain retained. The rejected raw snapshot was not saved by the old write ordering; it was not reconstructed. Commit446a85d moves its exclusive host write before the phase check and device mutation. An exact-warning regression proves retention and zero device mutations;45 focused tests and 5 subtests pass, with Ruff clean.

Under the owner’s existing temporary-disable authorization, Wi-Fi was again manually disconnected without saving configuration. A separate525.396-second quiet interval had exact overlap, no new alerts and unchanged system state. Reviewed snapshot SHA 18b56028320f38207ce4c85551f014d04f408698f8baf25582e685d90fe0b172 explicitly seeded b1 after byte-exact a1 archival. No automatic reseed occurred.

19e’s blocked host marker remains untouched. A new host evidence directory, 19e-wifi-quiet, copies26 immutable files from the original unqualified candidate, inventory SHA 0134a8115f38a5be9b615a419da262dbb72e75da6292116399c2d1102bca92a8. This supports19e’s first actual native Run; it is neither a native retry nor a repeated export. It passed with max activation error0.00048828125 and zero mismatches. Its old surrounding export interval remains failed.

19f’s first native export passed and produced232,384 bytes, NBG SHA 69eddc1ac45e2016f5adee2f53de6587fcc564eaaf90c4720e316b1d7e2b53f9. Raw parsing verifies CRCs and fixed FP16 I/O. Stock routing is24NN +1TP, with zero SH/CPU routes. The actual-size executor source/ELF passed independent review before the first full qualification. That speech-feature Run and three subsequent strict WAV frontend Runs passed. The frontend calls were wrapped in an ORT neural-session throw guard and each recorded one native call, zero CPU sessions and CLEAN cleanup.

Clip	Speech frames	FP32 decision agreement	Max speech-probability error
Public speech	175/200	100%	0.000814493
Silence	0/200	100%	0.000023429
Seeded noise	0/200	100%	0.000030908
Derived noisy speech	175/200	100%	0.000168334

All800 decisions agree with the original source-matched FP32 reconstruction; max native activation error is0.00390625 under the unchanged element policy. These are fixed-clip parity results, not a labeled VAD accuracy evaluation. The real speech clip uses exact frozen features for initial qualification; the other three additionally exercise the actual strict file frontend.

Paired benchmark and limits

One process ran36 CPU/NN pairs:4 warm-up and32 measured, alternating order. CPU is the original folded FP32 graph on stock ORT1.15.1 with one thread; native is the exact qualified FP16 NBG. Every native iteration reproduces the qualified output bytes. The CPU reference matches the earlier successful Wave baseline. Source, actual ELF, inputs, both references and preparation receipt were independently reviewed before the single benchmark attempt.

Prepared-feature measurement	CPU median / p95 ms	Native median / p95 ms
Run	8.577104 /9.252773	0.690792 /0.814966
Prepared-feature boundary	8.664375 /9.331614	0.833188 /0.993871
Calling-thread CPU time	8.545312 /8.760459	0.387897 /0.461373

Boundary median speedup is10.399069837×; calling-thread CPU time falls95.46071%. Native boundary includes input and NaN-canary copies, Run, output readback and softmax. CPU boundary includes input copy, Run and softmax; CPU output priming and both paths’ result validation are outside it. Feature extraction, FP32→FP16 feature conversion, session/graph setup and file/ADB orchestration are excluded. Caller-thread clocks do not measure total system CPU or energy. Every pair repeats the same frozen speech window. Each window is3.99seconds; the20ms output interval still requires about 1.46seconds of future context and does not establish causal streaming latency.

A host-only build attempt initially used relative Docker volume paths and was rejected before compiler/device execution. Its log is preserved; a new host directory with absolute paths produced byte-identical source/data. This consumed no native budget or retry. The final run count is6 compiler processes,12 export API calls and45 NPU Runs, within8/16/46 ceilings.

Final closure

Independent review reconstructs all35 b1 phase events and every cleanup advance from the reviewed seed through the benchmark. Every operation has empty health differences/alerts and CLEAN cleanup. Final direct seed-to-state diagnostic comparison also passes; kernel, boot, service identities, hashes, NPU settings and compiler absence are unchanged. All28 unique staging roots are absent and no galcore holders/native lab processes remain.

After closing the measurement phase, wlan0 was raised and reassociation acknowledged OK. At2026-10-14T10:01:43.929Z the original saved SSID/network ID were verified privately, supplicant COMPLETED, adminUP and an IP assigned. Credentials/configuration were not saved or changed. A separate post-restoration snapshot passes with no new kernel alerts or state deltas. The earlier blocked restoration attempts and all failed intervals remain retained.

Before the benchmark,13 altered artifacts of actual qualified19f copies refused with zero CPU sessions/transport and unchanged originals. After the benchmark, a separate pure-host check revalidates all123 prepared files and 15 packages, including12 qualified packages, again with zero CPU/transport. No production VAD service was installed. Further live work requires its own reviewed phase and bounded contract; no native retries are authorized by this completed experiment. Raw resumed evidence is retained under artifacts/npu/poc19/resume-20261014/; the tracked sanitized ledger contains the exact qualification, clip and benchmark receipts.