back to table · edit · history · where entries came from · files · download
3408 bytes, as of the version from 2026-11-16 22:11 (current). Recorded here, not run.
# T728 P1 repair evidence
P1 is done. The live reviewed revision 5272 (digest e0e0da8fe378294b28a6aca8f709940cfd8f8e31830dafd6070737184a284a59) reproduced the duplicated `mpmath` HTML ID: the rigour-details citation selected Programs, not Links. The only document change is `Programs.mpmath` -> `Programs.example`. Program language/code, Links.mpmath, mathematical prose, selection, all 105 values and original sources are retained byte-for-byte at the relevant field/file level. Document edit revision: 5273, digest 17aacfcfc5c8744c01df059fed3a38b5027161aa96c4676e5083c13a6a426d48.
Browser checks found one `mpmath` ID after repair, no duplicate IDs, correct click target `#mpmath` containing https://mpmath.org/doc/current/functions/zeta.html, exact program text and all 105 stored row texts. Known-good and deliberate duplicate-anchor controls pass. Raw authenticated page snapshots contain no API key; the browser uses saved HTML without a credential. Reproduce with `repair-presentation-check.py before` / `after`, setting TMPDIR to the downloaded inputs and using Playwright plus a Chromium installation.
`repair-checkpoints.py` freshly evaluates the gamma-envelope and pair arithmetic from the saved 107 full root/Hardy-jet input enclosures after verifying all numerical source/checkpoint hashes. All 105 resulting balls fit the unchanged stored decimal intervals; minimum retained accuracy is 503 bits. Saved 533/599-bit output balls and 130/165-dps mpmath values were also checked against every stored interval. `verify(sample=None)` using the explicitly labelled checkpoint replay returns 105/105 matched, zero differing/missing/extra. Known-value, perturbed-decimal, wrong-index, nonfinite and wide-ball controls pass. Launcher exit 0, Sage 10.9 / Python 3.12.3 / numberdb 0.1.10. No root or Hardy Taylor jet was rebuilt. This is reproducibility regression using unchanged builder inputs/formula, not a new independent mathematical verdict. See repair-checkpoint-results.json for all actual comparisons.
To reproduce the Sage check, download this draft's existing generate.py, hardy_arb.pyx, checks.py, check.py, argument-selection.md, skill-manifest.json, finite-prefix-roots.json, numerical-evidence.json and endpoint-checkpoints.json, plus before-snapshot.json, before-document.json and after-document.json. Pipe your own authorized key into `NUMBERDB_KEY_FROM_STDIN=1 sage -python repair-checkpoints.py` in that directory. The campaign itself ran this only through agents/sage.sh.
Canonical instructions: coordinated archive agents/runs/20261116T220407Z-repair-skills/manifest.json, bundle SHA-256 984965681e2f8925c511a121c118efcde59581c7b0d906cac8b641bf6925e55a. Complete local/live resource hashes match; read the local core and writing/submission/validation/computation references. See repair-skill-validation.json and repair-skill-manifest.json.
The post-edit audit retains only its original definition-length advisory; local presentation findings are empty. Programs participates in the selection fingerprint, so the earlier argument sidecar is stale even though selected indices and all generator bytes are unchanged. A fresh independent revision-and-attachments-bound scientific critique and argument review are required. No critique evidence or verdict was changed. No offering, publication or review action was performed. The controller owns subsequent validation and offering.