Map a model — and make the test able to fail

The hardware check includes deliberate invalid variants, not only success cases.

OUR OBSERVATIONLast verified: 2026-09-27 · Build 1.9.37Measured 2026-08-26 · BC.A1.003.009 · MetaTF 2.19.3

Quick reference. The long path with checkpoints: chapter 09 of the guide.

Measured mapping limits

Constraint	Observed limit
Input edge	≤ 256 px (“maximum input second dimension size is 256”)
Input layer	≤ 64 filters
Head input	≤ 256 channels (“Filter size is too big to fit in a CNP”)
Card	32 neural processors

Akida 1.0 consequences

The project uses InputConvolutional, Convolutional, SeparableConvolutional and FullyConnected. No generic upsampling/transposed-convolution path meant predicting rotated frames directly rather than masks.

V2–V6 and the negative controls

Variant	Input	Grid	Cell	Seq / passes	NPs	Result
V2	224 px	14×14	16 px	1 / 1	5	PASS (reference)
V3	160 px	10×10	16 px	1 / 1	4	PASS
V4	224 px	28×28	8 px	1 / 1	14	PASS
V5	256 px	16×16	16 px	1 / 1	5	PASS
V6	256 px	32×32	8 px	1 / 1	18	PASS — chosen
V2-M1 / V2-M2	224 px	14×14	16 px	1 / 1	5	PASS (M=1 / M=2 comparison heads)
X-gross	224 px	14×14	16 px	—	—	must FAIL (head input 512 channels)
X-px320	320 px	20×20	16 px	—	—	must FAIL (input > 256 px)

The variants are alternatives, not a sequence. Measured 2026-08-26, AKD1500 firmware BC.A1.003.009, MetaTF 2.19.3, random weights, hw_only=True, mode AllNps.

All V2–V6 variants pass one hardware sequence / one pass in the documented test. X-gross (>256 head channels) and X-px320 (>256 px input) must fail. A PASS is not trustworthy unless the deliberate negatives FAIL.

python3 tools/check_single_pass.py --list
python3 tools/check_single_pass.py

Expected: V2–V6 pass; V6 uses 18/32 NPs; both X variants fail.

Measured 2026-08-26: firmware BC.A1.003.009 · MetaTF 2.19.3. Mapping success is not an accuracy result.