30 — Repro for other AIs (shipping note)¶
Goal: Third parties (humans or other coding agents) cloning https://github.com/KanakMalpani/Binary-Neural-Networks get the same conclusions under published gates.
What was added¶
| Item | Purpose |
|---|---|
bnn/determinism.py |
Seeds + torch.use_deterministic_algorithms + CPU policy |
scripts/repro_all.py / bnn repro |
One-command verify / optional full smoke |
tests/golden_floors.json (v2) |
Floors for MNIST, image, audio, wrap, compression, native |
tests/test_golden_gates.py |
Pytest asserts vs committed results/*.json |
constraints.txt + pinned pyproject.toml |
Portable dep band (Python ≥3.11) |
REPRODUCIBILITY.md |
Human + AI runbook |
AGENTS.md |
Strict agent command order |
| CI updates | Windows + Linux run reproducible smokes |
Modes¶
- Fast verify: compile (best-effort) → pytest → export-check → validate-native (skip if no DLL) → golden compare → SUMMARY. No retrain.
- Full: + short deterministic smokes (default writes
_repro_smoke_*.json).
Guarantees¶
- Identical: compression 32×; native/NumPy GEMM err=0 (when path applies).
- Gated: accuracies within ±pp of published floors (not bit-identical floats).
Thesis lock unchanged. Datasets stay gitignored.