repair-scope.md

back to table · edit · history · where entries came from · files · download

3708 bytes, as of the version from 2026-11-16 12:25 (current). Recorded here, not run.

T724 repair run 20261116T104158Z, zeta3, codex-cli / gpt-6.1-sol / xhigh.

Confirmed and repaired only finding T724-P1: `Programs.Sage` is now a record containing `language: Sage` and `code`, with the original source string preserved exactly. All other sections, all 13 stored values, and all 26 original attachments are unchanged. The generator remains `generate.py`, SHA-256 `04c48e46d7c0be40da870b6211058f32f00277de6991332e93cc996df9081dde`.

The complete document was refreshed before the guarded PUT, which supplied X-Base-Revision for revision 5253 (`3d6bed425b31b08e3d7930e1e48d791f43da7523fa540759d6cc010a99651de9`). Repair produced revision 5254 (`85facff357dfc049f2274867648eaf7984e8a646411703ddefe787761b4ae989`). The bot's current engine/model/run provenance accompanied every write. Repair evidence attachments use source-hash preconditions and do not replace earlier evidence.

Numerical reuse is limited to unchanged mathematical inputs and generator bytes. `repair_numerical_check.py` replays the original exact binary-rational moment checkpoints, evaluates all 13 signed determinants, invokes Generator.verify(sample=None), verifies finite positive enclosures contained in the stored decimal intervals, and checks O3/O5 closed forms plus wrong-sign controls. It also executes the exact Programs example, n=7. All checks passed; the minimum retained accuracy is 376 bits. This replay is reproducibility evidence, not new independent quadrature. No unconditional whole-table integration was run, and no numerical string changed.

Reproduction: place generate.py, moments-checkpoint.json, after-document.json, and repair_numerical_check.py together with SageMath 10.9 and numberdb 0.1.10. Pipe your authorized key on stdin into `sage -python repair_numerical_check.py`. On this campaign host every Sage call goes through agents/sage.sh. The preserved log contains the actual REPAIR_NUMERICAL_JSON record; repair-numerical-results.json is its parsed copy.

The live rendering check, repair_render_check.py, uses the saved before/after API documents and Playwright 1.64.1 / Chromium 156.0.8078.4. Run it with TMPDIR set to their directory and the documented shared browser PYTHONPATH, LD_LIBRARY_PATH, PLAYWRIGHT_BROWSERS_PATH, and cached FONTCONFIG_FILE. Eleven bounded section views include the program, all values, definition, comments, formulas/references, related tables, links, coverage and all expanded rigour paragraphs. All render without parse, MathJax or unresolved-markup errors. The canonical private page still needs a website session cookie, so these are same-renderer section checks, not a whole canonical-page inspection. Screenshots and HTML are retained in the run directory.

The archived canonical skill manifest is agents/runs/20261116T104158Z-repair-skills/manifest.json, SHA-256 c5ee30e128246f511fb7c0e8e37f8b7d2419d5fad25227688a6111fdf9686049. Both local/live bundles and all their archived resources were hash-validated and match bundle SHA-256 d6661c2f21613524889fa3a57d3c95f8330344ca62e5a856211c08a282b7dbab. Read the local core plus writing, submission, computation and validation references only. No argument selection or domain was changed.

The Programs container participates in the selection fingerprint, which changed from ea2dc56c655ee6d27d5201656002b848904096774f47dd1e67b9f378e41cef54 to cb04d6cd9f2ef63d7d7158e504b7b383af0518ab6a371b74b66024b8edd23812. A fresh independent scientific verdict and argument sidecar are required for the exact final revision and attachments. The old critic JSON, Markdown, evidence and sidecar are immutable and were not edited. This repair supplies no independent acceptance, offers nothing for review, and publishes nothing.