Validation report / 验证报告
PASS_LOCAL_SMOKE. Historical local smoke record for the public research preview; license Apache-2.0 (scoped).
Source: 3ffdf3d0283689d3b0da379b04dd0304d30c75df
Fresh environment: Python 3.11.6, torch 2.10.0+cpu, safetensors 0.7.0; all resolved versions pinned in requirements.lock.txt. pip check passed.
Command: python evaluate.py <trusted-local-checkpoint>; source hash verified before loading.
Verified 7524640 parameters. Converted tensors match every hash-pinned original checkpoint tensor exactly.
| Check / 检查 | Result |
|---|---|
| safe_load | PASS |
| normal_ranking | PASS |
| raw_unknown_NOUL | PASS |
| host_unknown_refusal | PASS |
| all_invalid_host_refusal | PASS |
| blocked_candidate_invariance_exact | PASS |
| manifest_integrity | PASS |
| trusted_checkpoint_tensor_parity | PASS |
Normal selection: quality. Raw unknown NOUL logit: 1.6348905563354492. Unknown/all-invalid host refusals did not call the model. Blocked-candidate outputs matched exactly.
CPU one thread, 50 calls: median 2.751 ms; maximum 3.638 ms. Single-laptop smoke, not production performance.
Provenance and license follow-up
Apache-2.0 is selected for owner-held rights in this independent package. Trainer, both generators, manifest, all five splits and checkpoint hashes match dated 2026-09-26 training evidence. Historical metrics remain unreproduced. Historical model implementation hash mismatch and missing training Git revision remain unresolved. This independent public release has been explicitly authorized. See PROVENANCE_AUDIT.md and LICENSE_SCOPE.md. No new training or heavy evaluation was run in this follow-up. Existing weights are unchanged.