Findings · Assurance architecture
A training run's final loss and trajectory re-derive bit-for-bit from its recorded inputs on one machine; cross-hardware re-derivation is left as UNKNOWN, and not claimed.
Evidence
The verifier does not trust the recorded result: it re-runs training from the recorded configuration and re-derives the trajectory and final loss, and reports VERIFIED only when they match bit-for-bit under a matching numeric environment. The interesting part is the boundary. A re-derivation mismatch under a different environment returns UNKNOWN, never TAMPERED, even a genuine forgery of the loss is only reported as UNKNOWN there. Detection power degrades exactly where reproducibility does, and the tool says so instead of overclaiming.
Sample. One machine (Apple M4 Max, NumPy on Apple Accelerate), a handful of runs, plus a 12-case negative-control suite in which each verdict state is triggered by a real tamper.
Boundary
Bit-for-bit reproducibility on other hardware, which is untested and reported UNKNOWN rather than upgraded. It does not establish code correctness, buggy-but-faithful code still verifies, nor result quality nor accountability.
Independent reproduction. None known. The run was re-derived bit-identically from a fresh public clone during the evidence inventory, which is author self-reproduction, not independent.
Reproduction
./run_all.sh- Toolchain
- Python 3.14.4, NumPy 2.4.4 on Apple Accelerate BLAS
- Expected output
- VERDICT VERIFIED; final_val_loss 0.10635284871660042; compute_fingerprint c6da865de1f953d9aa4c1e4738b99a8c; 12 of 12 controls pass
- Claim stated at
- repowazdogz-droid/nanogpt-provenance · RESULTS.md; README.md:50-51; GAPS.md:56-61
- Verified at
- repowazdogz-droid/nanogpt-provenance@732da8b1 (2026-07-20)
This pass. Re-verified this pass: run_all.sh was re-run from a fresh clone at 732da8b1, reproducing the loss and fingerprint bit-for-bit (12 of 12 controls).
Artifact: repowazdogz-droid/nanogpt-provenance