PoC 14A — Learned Affine-plus-Erf Host Contract

Date: 2026-07-29 Device calls: 0 Result: PASS for the host model; deliberately unqualified for compilation or execution

Question

Can the CPU-resident learned affine stage from PoC 13 be frozen as a finite, integer-defined Multiply -> Add -> Erf partition without losing the noise suppression quality gates?

PoC 14A does not ask whether the Wave compiler accepts the partition. It freezes the mathematical target, its quantizer, its test vectors, and its package before a native compiler exists.

Frozen partition

feature: UINT8 [481], scale 0.0625, zero point 128
  -> per-bin Multiply by INT8 [481]
  -> Add INT32 bias [481]
  -> requantize to UINT8, scale 0.03125, zero point 128
  -> Erf
  -> UINT8 [481], scale 0.0078125, zero point 128

The selected quantization candidate is:

name:               u8-per-channel
feature:            UINT8 affine-asymmetric
weights:            INT8 symmetric-per-channel
bias:               INT32, feature_scale * weight_scale
rounding:           nearest, ties to even
saturation:         explicit for feature and Erf input
contract SHA-256:   b52ddaf3baed2a7001ff10cbf9ba0600945749a2332f02d71d2ae0be25ac98b8
target product:     0x1000000e

The exact integer oracle centers the feature, multiplies by the quantized weight, adds bias in accumulator units, requantizes with frozen float32 multipliers, and indexes the already qualified 256-entry U8 Erf table.

Finite candidate registry

The host compared exactly four candidates:

u8-per-tensor
u8-per-channel
i8-per-tensor
i8-per-channel

All use feature scale 0.0625; U8 uses zero point 128 and I8 uses zero point 0. Weights are always INT8. The registry contains no dynamic type, scale, zero-point, rounding, or saturation choice at inference time.

Validation used 16,384 held-out 481-bin frames. Selection was minimum MSE, then maximum error, then registry order:

Candidate	Exact fraction	Max U8 error	MAE	MSE
U8 per tensor	0.9799457003	5	0.05120444	0.18419382
U8 per channel	0.9814606665	5	0.04615527	0.16474721
I8 per tensor	0.9799457003	5	0.05120444	0.18419382
I8 per channel	0.9814606665	5	0.04615527	0.16474721

The independent 32,768-frame test split was confirmation-only:

Candidate	Exact fraction	Max U8 error	MAE	MSE
U8 per tensor	0.9802865328	5	0.05047354	0.18160852
U8 per channel	0.9818325241	5	0.04532660	0.16185984
I8 per tensor	0.9802865328	5	0.05047354	0.18160852
I8 per channel	0.9818325241	5	0.04532660	0.16185984

U8 per-channel wins by registry order after the U8/I8 numerical tie.

Six frozen vectors

The qualification bundle contains exactly:

1. zero point;
2. minimum;
3. maximum;
4. repeated ramp;
5. seeded random, seed 346387253;
6. learned validation frame 5,513.

The binary bundle contains six inputs and six independent oracle outputs. Its manifest is bound to the model contract and records each tensor digest.

affine-vectors.bin SHA-256:
  fca16527e4efec7f726ceecad1e5d704dff1dd6658ef5062c71900a4948981dd

affine-vectors.json SHA-256:
  5c8b6184c6b2acba4cc6dd82e13df93eecab573de27e915739c7fccb20a4653b

Full held-out evaluation

The new feature-boundary partition was injected into the existing causal DFN-or-Erf selector. No STFT, noise estimator, synthesis, DFN output, corpus, metric, or selector implementation was forked.

The run evaluated 1,230 items and rejected none:

Gate	Result	Threshold	Status
Synthetic median SI-SDR improvement	+3.256976 dB	at least +3 dB	PASS
Natural median SI-SDR improvement	+5.233915 dB	at least +1 dB	PASS
Natural median STOI delta	+0.013537	at least -0.02	PASS
Noise-only median attenuation	10.583481 dB	at least 6 dB	PASS
Clean-speech median absolute RMS delta	0.046372 dB	at most 1 dB	PASS

The evaluation made 740,742 partition calls. The selector chose the learned Erf candidate 450,140 times and the CPU DeepFilterNet candidate 290,602 times.

The complete evaluation is retained at:

artifacts/npu/runs/poc14a-erf-affine-host-20260729/evaluation.json
SHA-256:
  ee1c5cedbf0a1ab63f311ff07ca3b29b842673d395a90b5d0b9df26a0ec94647

Deterministic package

Two independent builds produced byte-identical copies of all nine files:

artifacts/npu/runs/poc14a-erf-affine-host-20260729
artifacts/npu/runs/poc14a-erf-affine-host-20260729-regen

The first build used the independently reproduced DFN v3-a evidence; the second used v3-b. Both DFN payloads were already byte-identical. Package hashes:

Artifact	Bytes	SHA-256
affine-config.json	1,543	a8cdec21a3bee0b2c29b70bcd4adfe0f2fc5beffd9fdd3002b3af12d6800a536
affine-model.onnx	5,983	3e8658d482626289443085d8cbeabc0f1a7130b1d5f00cbf8884ee6fb2478a0a
affine-vectors.bin	5,824	fca16527e4efec7f726ceecad1e5d704dff1dd6658ef5062c71900a4948981dd
affine-vectors.json	2,469	5c8b6184c6b2acba4cc6dd82e13df93eecab573de27e915739c7fccb20a4653b
affine-weights.npz	7,211	9e26df0a85570f99add70268a3e7a76dafd49bac263377602de4be490414855a
erf-u8-lut.bin	256	c9525963e022d4a4aea0814845119b37488f3727ec8d96684f52e1e2ad776296
evaluation.json	2,042,774	ee1c5cedbf0a1ab63f311ff07ca3b29b842673d395a90b5d0b9df26a0ec94647
source-model.json	2,938	15096a6ffb680bae6285badc840791be14279b5c50448ba4278a8427ffe5507a
model-manifest.json	3,666	99cccb924ab807171ce93b5cba40b3d7ed52219294c0f5d3104b153aa0294263

Strict reload reconstructs the quantized model and requires canonical byte-for-byte reserialization. The ONNX model passes checker validation and ONNX Runtime agrees with its exact reference graph.

The manifest intentionally remains:

{
  "compiler_qualified": false,
  "execution_approved": false,
  "nbg": null,
  "numerical_evidence_sha256": null
}

Host quality is not compiler or NPU evidence.

Native constants boundary

The selected model also has one fixed little-endian native representation:

magic:               AFC1
schema:              1
width:               481
input type:          UINT8
weight granularity:  per-channel
header bytes:        68
payload layout:
  INT8 weights             481 bytes
  float32 weight scales  1,924 bytes
  INT32 biases           1,924 bytes
  float32 accumulator    1,924 bytes
total:                 6,321 bytes

The header embeds the 32-byte model-contract digest. The parser validates every scalar, vector shape and type, reconstructs the model, recalculates the contract, and requires canonical byte identity.

artifacts/npu/build/poc14b/affine-constants.bin
SHA-256:
  a2b588be1a4e68a3179a18b51be016115782f923344521c667ca4310c9eeecb4

Reproduction

From the repository root:

uv run --project waved/tools/onnx-hybrid wave-erf-affine build-host \
  --model artifacts/npu/runs/poc13a-erf-mask-host-20260729-model-v4-selector \
  --examples artifacts/npu/runs/poc13a-erf-mask-host-20260729-examples-v2 \
  --source-evaluation artifacts/npu/runs/poc13a-erf-mask-host-20260729-evaluation-v5-selector-final/evaluation-summary.json \
  --slr31 artifacts/npu/corpora/slr31-peak-v1 \
  --stream artifacts/npu/runs/poc13a-erf-mask-host-20260729-evaluation-stream-v3/evaluation-input.f32 \
  --dfn-output artifacts/npu/runs/poc13a-erf-mask-host-20260729-dfn-output-v3-a.f32 \
  --dfn-run artifacts/npu/runs/poc13a-erf-mask-host-20260729-dfn-run-v3-a.json \
  --output artifacts/npu/runs/poc14a-erf-affine-host-20260729

uv run --project waved/tools/onnx-hybrid wave-erf-affine verify-host \
  --package artifacts/npu/runs/poc14a-erf-affine-host-20260729

uv run --project waved/tools/onnx-hybrid wave-erf-affine \
  export-native-constants \
  --package artifacts/npu/runs/poc14a-erf-affine-host-20260729 \
  --output artifacts/npu/build/poc14b/affine-constants.bin

The output paths are non-overwriting. Use a new path for regeneration.

Decision

PoC 14A passed. It proves a useful, deterministic host partition and CPU fallback contract. It does not approve a graph for the Wave. PoC 14B must independently show that the exact Wave producer accepts the frozen tensor types before any NBG differential can be interpreted.