Novel paper candidates (local research vault)¶
Date: 2026-08-15
Policy: Selective — only credible novel systems contributions from this lab; not XNOR-Net/BitNet surveys.
Tasks: W12.T01 (vault links) · W12.T05 (triage → ship or defer) · W12 B1 skeleton (in-repo)
Local vault links (W12.T01)¶
| Resource | Path |
|---|---|
| Vault root | C:\00 Research Papers |
| Series index | C:\00 Research Papers\BINARY_NEURAL_SERIES_README.md |
| B1 folder | C:\00 Research Papers\Stop Claiming 32x Honest Speedup Accounting for Binary Neural Networks\ |
| B2 folder | C:\00 Research Papers\Packed XNOR on Commodity CPUs Reproducible Productization and Golden Gates\ |
| B3 folder | C:\00 Research Papers\When Not to Binarize Decision Tree for Hybrid Low-Bit Wrapping\ |
Each folder contains: README.md, 01_idea.md, 00_source/ (SOURCES.md + extracted notes).
Lab publication plan + claims whitelist: docs/PUBLICATION_PLAN.md.
B1 in-repo skeleton: docs/papers/B1_STOP_CLAIMING_32X.md.
Figure / claims machine check: bnn bridge figures.
Created folders (3)¶
| Paper | Path | Novel claim (1 sentence) |
|---|---|---|
| B1 — Honest speedup accounting | C:\00 Research Papers\Stop Claiming 32x Honest Speedup Accounting for Binary Neural Networks\ |
Dual-metric reporting (op-count/compression vs wall-clock) plus fake-binary negative control and Amdahl/energy proxies — never advertise 32×/64× as latency. |
| B2 — Packed XNOR productization | C:\00 Research Papers\Packed XNOR on Commodity CPUs Reproducible Productization and Golden Gates\ |
Commodity CPU packed XNOR recipe (MSVC/portable) with fair pre-pack benches and machine-checkable golden gates for third parties/AI agents. |
| B3 — When not to binarize | C:\00 Research Papers\When Not to Binarize Decision Tree for Hybrid Low-Bit Wrapping\ |
Goal×hardware decision tree + hybrid FFN wrap evidence for honest wrapping vs INT4/FP8/GGUF / “do not binarize.” |
Lab evidence mapped into papers¶
| Evidence | Primary paper |
|---|---|
docs/06, results/benchmark.json, fake-binary trap |
B1 |
results/energy_bound.json, (E=P\cdot t) proxy |
B1 |
docs/35, bnn/math/ XNOR↔dot proofs + Amdahl calculators |
B1 (math spine) |
results/math_ste_compare.json (STE / ApproxSign / EDE) |
B1 (learning math) |
AGENTS.md, docs/30, tests/golden_floors.json, bnn repro |
B2 |
| Multimodal canaries (MNIST / CIFAR / audio synth) | B2 (canaries, not SOTA) |
docs/18 decision tree, docs/12 wrapper taxonomy |
B3 |
results/hybrid_ffn_wrap.json, wrap_demo.json |
B3 |
docs/20 NPU vendor closure |
B3 |
docs/23–24, bnn bridge … |
B3 (bridge honesty) |
Triage → ship or defer (W12.T05)¶
| Candidate | Decision | Rationale |
|---|---|---|
| B1 Honest speedup accounting | SHIP (tech-report primary) | Unique dual-metric + fake-binary control; evidence already in committed results/benchmark.json. In-repo skeleton: docs/papers/B1_STOP_CLAIMING_32X.md. arXiv / PwC remain human. |
| B2 Packed XNOR productization | SHIP (companion / systems note) | Matches repo thesis + bnn repro / golden floors; citable artifact path |
| B3 When not to binarize | SHIP (short companion) | Decision tree + bridges CLI; prevents false BNN wins vs INT4/bitnet.cpp |
| Literature-only BNN survey | DEFER permanently | No novel systems claim |
| Cross-modality “SOTA lab” | DEFER permanently | Proxies only; forbidden as golden |
| Standalone energy methodology | DEFER (folded into B1) | Insufficient alone |
GPU 32× from sign() |
REJECT | Thesis lock |
Drafting must preserve thesis lock: packed CPU/edge kernels + honest STE simulation; no fake GPU wins.
Rejected standalone angles¶
- Literature-only survey of BNNs
- Cross-modality “SOTA lab” from proxy CIFAR + synthetic audio
- Energy methodology alone (folded into B1)
- Any GPU 32× claim from
sign()(forbidden by thesis lock)
Status¶
| Paper | In-repo | Next (human) |
|---|---|---|
| B1 | Markdown skeleton docs/papers/B1_STOP_CLAIMING_32X.md — C1–C7 only |
arXiv submit + Papers with Code official-code registration |
| B2 | Vault + triage only | Companion after B1 |
| B3 | Vault + triage only | Companion after B1 |
Vault linked + triage closed for v1.0 narrative. Venue LaTeX still optional; claims must pass bnn bridge figures against goldens. This PR does not submit arXiv.