Publication plan (W12.T02)¶

Field	Value
Status	In-repo B1 skeleton (not submitted; arXiv is human)
Date	2026-08-15
Venue (candidates)	Tech report on GitHub + optional workshop (MLSys / edge-AI)
B1 skeleton	`docs/papers/B1_STOP_CLAIMING_32X.md`
Paper vault	`C:\00 Research Papers` — see `docs/32_NOVEL_PAPER_CANDIDATES.md`
Papers with Code	Official-code URL ready; register after arXiv (human)

Claims whitelist (must match goldens)¶

Allowed to claim in any write-up — each row ties to tests/golden_floors.json and/or committed results/*.json:

ID	Claim	Evidence gate
C1	Aligned uint64 binary pack compression 32.00× (theory, not latency)	`compression_exact_when_uint64_pack` + `results/wrap_demo.json` / benchmark theoretical
C2	Native XNOR-popcount GEMM err = 0 vs ±1 FP when DLL/`.so` present	`native_err_max: 0` + `results/benchmark.json`
C3	Dual-metric culture: theory vs wall-clock; never GPU 32× from `sign()`	Thesis lock + floors `notes`
C4	MNIST binary MLP within floors	`mnist.binary_mlp_min_acc` vs `results/train_results.json`
C5	CIFAR Bi-Real proxy within floors	`image_cifar.binary_bireal_min_acc` vs `results/image_cifar.json`
C6	Audio synth binary CNN within floors	`audio_synth.binary_cnn_min_acc` vs `results/audio_synth.json`
C7	Linux + Windows CI; Linux native `.so` validated in Actions	CI matrix / WC-R2

Machine-check the whitelist:

bnn bridge figures --out results/figures_manifest.json

Forbidden:

GPU e2e 32× from STE/sign()
Invented bench shapes as “the” golden
Bit-identical floats across machines as a pass criterion
Production ASR / full ImageNet SOTA as delivered

Figure pipeline (W12.T03)¶

bnn bridge figures --plot-dir results/figures
bnn pareto --from-results --out results/pareto_from_results.json --plot results/pareto_from_results.png
python scripts/pareto_report.py --demo --out results/pareto_demo.json --plot results/pareto_demo.png

Prefer figures generated from committed results/*.json + Pareto / figures-manifest schemas. Manual polish OK; source JSON must stay in repo. No invented goldens.

Bridges (CLI)¶

bnn bridge list
bnn bridge gpu --probe
bnn bridge cpu-llm

See docs/23_BITNET_CPP_BRIDGE.md and docs/24_GPU_INT4_FP8_LANE.md.

Papers with Code (code link)¶

Official implementation (no arXiv ID yet): https://github.com/KanakMalpani/Binary-Neural-Networks

Field	Value
Title	Stop Claiming 32×: Honest Speedup Accounting for Binary Neural Networks
Skeleton	`docs/papers/B1_STOP_CLAIMING_32X.md`
Citation	root `CITATION.cff` (`bnn-lab` 1.0.0)
Tasks	Quantization / model compression — not ImageNet SOTA, not ASR product
Human	After arXiv: add paper by ID, mark this repo official code

Do not upload to arXiv or click PwC submit from CI / this PR.

Citation¶

See root CITATION.cff (version aligned to release tag). Cite the B1 skeleton path until an arXiv ID exists.