| # Validation | |
| Status: **PASSED** for the replay scope; publication remains **PAUSED / PRIVATE REVIEW**. | |
| ## Pre-release clean-room replay | |
| The pre-release validation began from a target declared absent by the admitted operation contract, fetched and SHA-verified all 36 public source archives, and ran 51 tests (49 derivation tests plus two boundary tests). It then generated all 49 scientific payloads before reference comparison. Every payload matched, and all source-corruption, result-mutation, and expected-answer-isolation falsifiers passed. | |
| `validation/clean_room_replay_receipt.json` binds these results to the exact replay/test/reference/source-lock file tree in this candidate. The receipt itself is provenance evidence; evaluators should still run the commands in `REPLAY_STANDARD.md` because receipt checking alone is not scientific replay. | |
| ## Independent private-Hub download replay | |
| Ouroboros also exercised the evaluator path from the uploaded private Hugging Face repository. It downloaded the pinned commit into a newly absent destination, created a new virtual environment from `requirements.txt`, began with an empty source cache, fetched and SHA-verified 36 of 36 public source archives, passed 51 of 51 tests, generated 49 fresh scientific payloads, and matched all 49 only after generation. Source corruption, result mutation, and expected-answer-isolation falsifiers all passed. | |
| The first external validation attempt downloaded the complete repository but stopped in the validator because it expected a manifest field named `size_bytes`; the release manifest correctly uses `byte_count`. The downloaded repository itself had not failed. Ouroboros corrected only the external validator and repeated the entire download into a second absent destination. That repeat passed every inventory, manifest, source, test, compute, comparison, and falsifier gate. | |
| As an additional isolation test beyond the documented commands, computation was rerun while runtime access to the entire `references/` tree was forcibly denied. All 49 payloads were generated with zero reference-file opens. They were byte-identical to the normal fresh outputs and matched all 49 reference assertions afterward. | |
| These results validate scientific replay from public inputs. They do not replay the preceding autonomous research campaign. | |