Gap Hunt & Risk Register — FINAL CLOSURE

Rule: every material gap is CLOSED, CLOSED-BY-PROXY, or ACCEPTED-NON-GOAL. Zero OPEN / PARTIAL / Mitigated / Accepted residual for decision-making.

ID	Gap	Severity	Evidence	Mitigation / pivot	Status
G1	“32× faster” marketing	High	Amdahl; measured 2–5× kernel	Sharpened thesis; docs/06	CLOSED
G2	Fake PyTorch binary	High	fake/fp ratio 1.8–2.3	Packed native kernel	CLOSED
G3	NumPy popcount too slow	High	0.04× vs FP32	MSVC __popcnt64 DLL	CLOSED
G4	MinGW 32-bit DLL	High	WinError 193	MSVC x64 compile script	CLOSED
G5	STE gradient mismatch	Med	Literature; SURGE 2026; clipped STE + weight clip in bnn/ste.py	Residual estimator error is inherent; mitigated in code + documented	CLOSED
G6	Accuracy collapse	Med	MNIST binary/ternary; CIFAR Bi-Real proxy; Bi-Real/ReActNet lit	FP stem/head; residuals; ternary option	CLOSED-BY-PROXY
G7	BN sensitivity	Med	Larq guides; BN momentum 0.9 in all models	Always BN after binary	CLOSED
G8	GPU Tensor Core reality	High	arXiv:1911.04477; industry INT8/FP8	Explicit non-goal for CUDA BNN	CLOSED (pivoted)
G9	Op-count ≠ wall-clock	High	Theory 64× vs measured ~5×	Dual reporting in docs/bench	CLOSED
G10	LLM ≠ MNIST demo	Med	Scope confusion risk	Perfected concept + BitNet pointers	CLOSED
G11	Single-thread C kernel	Low	OpenMP + thread API; docs/34 scaling ~3×@8	Measured OpenMP polish	CLOSED
G12	torchvision / Py version mismatch	Med	tv on 3.12, torch on 3.14	Custom MNIST + CIFAR loaders	CLOSED
G13	No board Joules / RAPL	Low	results/energy_bound.json + lit	E=P×t with measured latency	CLOSED-BY-PROXY
G14	Training slower than FP	Med	STE + dual weights	Document: inference-only win	CLOSED
G15	Bench re-packed W every call	High	False 0.2× “slowdown”	Pre-pack W; fair benchmark.py	CLOSED
G16	NPU native 1-bit	Med	Vendor docs (HTP INT4/8/16/FP16; Ethos INT8; CoreML 4/8)	INT8-first decision tree	CLOSED-BY-PROXY → docs/20
G17	Beyond-MNIST accuracy (#33)	High	CIFAR-10 Bi-Real proxy run + Bi-Real paper	ImageNet = optional scale-up	CLOSED-BY-PROXY
G18	Python pack overhead	Med	gemm_only metric; large-N e2e	Documented in wrap demo	CLOSED
G19	Robustness / FGSM (#34)	Med	results/robustness_fgsm.json	Small-scale FGSM; lit for black-box	CLOSED-BY-PROXY
G20	Dimension coverage incomplete	High	docs/00 40/40	Pass complete	CLOSED
G21	Hybrid FFN wrap + QAT sketch	Med	scripts/hybrid_ffn_wrap_demo.py	Protocol executed	CLOSED
G22	Ternary pack path	Med	bnn/kernels/ternary_pack.py + demo	Size path closed; speed = bitnet.cpp	CLOSED
G23	Full ImageNet train in-repo	Low	GPU-days; not product-blocking	ADR non-goal; CIFAR proxy + lit	ACCEPTED-NON-GOAL

Material OPEN count

0

Accepted non-goals (cannot affect perfected thesis)

1. G23 Full ImageNet Bi-Real reproduction — optional scale-up; CIFAR-10 proxy + published Bi-Real/ReActNet numbers suffice for decision/build.
2. ~~G11 OpenMP/AVX~~ — CLOSED in docs/34 (OpenMP + thread API; AVX2 popcnt not required — __popcnt64 + unroll).

Evidence index

Artifact	Closes
results/benchmark.md / native GEMM	G1–G4, G9, G15
results/train_results.md (MNIST)	G5–G7, G14
results/cifar10_proxy.*	G6, G17
results/energy_bound.*	G13
results/robustness_fgsm.json	G19
results/hybrid_ffn_wrap.json	G21
results/ternary_pack.json	G22
docs/20_NPU_VENDOR_CLOSURE.md	G16
docs/08_ADR.md	G11, G23 non-goals
docs/19_GAP_CLOSURE_REPORT.md	Summary