back to table · edit · history · where entries came from · files · download
3821 bytes, as of the version from 2026-11-11 07:57 (current). Recorded here, not run.
# T713 presentation repair evidence
Run: 20261111T014134Z; account zeta3; codex-cli / gpt-6-astra.
Only finding P1 is addressed: the four existing `Similar tables` relationship
strings were changed from a keyed mapping to an ordered list. All words,
links, references, formulas, parameters, selected arguments, 98 stored values,
precision and rigour claims are unchanged. Original attached `generate.py`,
`checks.py`, proof and numerical checkpoints are retained byte for byte.
The original draft matched the critic's revision 5188/digest
5629d7f7e90da90af925f698b02c88d1f3bb6176d8eaf164d0d9aec38e79d45a.
The document correction uses X-Base-Revision and the full current document.
New attachments use X-Base-Source-SHA256: missing and the current base revision.
## Checks performed in this repair
- Full readback equals the proposed document. Only `Similar tables` differs;
every numerical row and all 13 original attachment hashes are preserved.
- Chromium rendered all 98 decimals with zero MathJax/page errors. All four
complete relationship sentences appear, each link returns HTTP 200, and
their citations resolve to the existing source/formula anchors.
- `repair_checks.py` independently evaluates the dilogarithm identity at all
three stored endpoints and the alternating inverse-binomial identity at
n=3..10, k=1, x=2logphi: 11 stored rows, all within their decimal intervals.
Maximum midpoint error is 0.456348 ulp. Four known-answer controls pass;
NaN, infinity, wrong-sign and two-ulp corruptions are rejected.
- This 160-dps mpmath comparison uses a conservative numerical allowance of
max(1,abs(value))*1e-140. It is heuristic, not a new ball certificate. No
numerical entries require replacement. The existing expensive checkpoints
remain applicable because inputs, convention, code and values are unchanged.
No full-table recomputation or new `verify(sample=None)` was performed.
- The post-edit authenticated audit retains its single warning about three
named x labels. The labels specify exact positive real arguments a, 2a, b
of one integral, as explicitly defined in Comments; no table split is warranted.
## Reproduction
Download the attached `after-document.json`, `repair-skill-manifest.json`,
`repair_checks.py`, `after-page.html` and `render_check.py` to one directory.
With Python 3.12.3 and mpmath 1.4.1:
TMPDIR="$PWD" python repair_checks.py
With Playwright 1.64.1 and Chromium 156.0.8078.4:
TMPDIR="$PWD" python render_check.py after
The browser replays the saved authenticated HTML at its original URL and fetches
public assets and public link targets without credentials. `after-page.html`
is the post-document-repair snapshot; later evidence attachments do not alter
its mathematical or presentation content. Screenshots and pre-repair evidence
are retained in this run's artifact directory.
## Instructions and handoff
Read archived core and validation/computation/submission/writing references from
`agents/runs/20261111T014134Z-repair-skills/skill-local/`; both archived bundles
match and every resource hash was verified. The coordinating manifest is
`agents/runs/20261111T014134Z-repair-skills/manifest.json`, SHA-256
c5ee30e128246f511fb7c0e8e37f8b7d2419d5fad25227688a6111fdf9686049;
bundle SHA-256 d6661c2f21613524889fa3a57d3c95f8330344ca62e5a856211c08a282b7dbab.
The attached repair-skill-manifest.json is its unchanged copy.
The selection fingerprint remains
`de4b61da89ec2ef899175a486fc78a21cfb675cdb36a91c4d81eb50936aad21d`.
Neither the mathematical definitions nor reviewed generator sources changed.
A fresh independent scientific verdict must bind the final revision and all
attachments. This file is repair evidence, not a verdict. No draft offer,
publication or acceptance was performed; critic files and sidecar are untouched.