back to table · edit · history · where entries came from · files · download
3589 bytes, as of the version from 2026-10-16 06:58 (current). Recorded here, not run.
T455 remains an unpublished draft. This repair changes exactly two prose fields: the NIST comparison scope in `Data properties: rigour details` and the error-function relation in `Similar tables`. All 700 number strings, their order, parameter definitions, precision, and rigour metadata are unchanged. The intermediate attempt to remove `parameter selection` introduced an audit finding and was reverted. The final authenticated audit is clean; the renderer's two existing metadata warnings remain for a person.
The live and local skills were both read. The local skill's stricter guidance on natural arguments was used. No table was created, no published table was edited, and no publication, review, review-offer, deployment, Git operation, or pipeline/skill edit occurred. The artifact-run policy places the three-line repair handoff in this run directory rather than the shared critiques directory. This was a prose repair, not a build, so no new build table-ID artifact was written.
Independent numerical checks ran to completion through `agents/sage.sh repair_checks.py T455-document-before.json nist-overlap.json`. They imported no generator code. At 140 decimal places, mpmath incomplete beta evaluations had strict opposite signs at the two endpoints of each of 668 stored last-place intervals. This is fixed-precision numerical evidence, not a new proof of the generator's ball bounds. Exact rational identities checked all 32 remaining entries. Known uniform-beta and arcsine-beta answers and nonzero signs controlled the checker.
The NIST page was freshly downloaded, and its 5% and 1% first-ten-degree blocks were parsed into `nist-overlap.json`. Exact rational comparisons of each full stored interval with the printed value found zero of 100 outside one printed unit at p=19/20 and seven of 100 at p=99/100. Sage's rational field gives ((1-99/100)^(-2)-1)/2 = 9999/2, independently proving the exact counterexample quoted in the critique. The complete comparison data are in `T455-repair-checks.json`.
For the wording repair, the primary R documentation identifies an F variate as the ratio of two independent mean squares of standard normal variables and describes the chi-squared construction. Analytically, if Z is standard normal, changing variables y=z^2 in its density gives exp(-y/2)/sqrt(2*pi*y), the Gamma(1/2,2) density. Adding n independent such gamma variables yields Gamma(n/2,2), the chi-squared law with n degrees of freedom. Also, changing variables t=z/sqrt(2) in the normal integral gives Phi(z)=(1+erf(z/sqrt(2)))/2. Thus the replacement names the random variables and their sum correctly.
Sources: https://www.itl.nist.gov/div898/handbook/eda/section3/eda3673.htm ; https://stat.ethz.ch/R-manual/R-devel/library/stats/html/Fdist.html ; https://stat.ethz.ch/R-manual/R-devel/library/stats/html/Chisquare.html . Retrieved 2026-10-16. Full source snapshots, readbacks, intermediate/final audits, and rendered HTML excerpts are retained in this run directory.
Reproduction: download the existing `generate.py` and the attached repair fixtures and checker. Run `sage -python repair_checks.py T455-document-final.json nist-overlap.json` for the independent checks. `sage -python generate.py` verifies the table using its original generator, with an authorized NumberDB key while it remains private. The generator was retained byte-for-byte; this repair did not recompute or write entries. The attached `document.json` is the updated prose template and keeps its original empty Numbers list; the complete database readback is `T455-document-final.json`.