By @rpinho

AI for physics: certified bounds on published results, exact checks with independent referees, and bug reports on physics software

6dfea25 / friends/penedones-de-sitter-bootstrap/VERDICT.md

Raw
HOLDS WITH CAVEATS
Referee: gpt-6.1-sol

The new bound survives: I-tilde_(2,0) <= 34113/6250 = 5.45808, with the paper's same normalization, external dimension 5/8 and principal-series gap 8.53.
[The primary paper](https://arxiv.org/abs/2107.13871), equation (5.53), really prints 5.67049. Its authors include João Penedones, EPFL, alongside Hogervorst and Salehi Vaziri.

I wrote two evaluators from the paper's definitions, importing neither claimant search nor verifier code. Direct gamma/3F2 evaluation reproduces the specified baseline 6.43173726057 and the new objective 5.4580734714720031025.
My separate Arb endpoint-series certificate proves positivity on all 61 closed intervals covering nu=8.53 through 39, every remaining even discrete mode, and both infinite tails. It also proves the upward rounding to 5.45808. See referee_certificate/VERIFIED.json, DERIVATION.md and REFEREE_EVIDENCE.tar.gz.

The trivial controls do not explain away the gain: the single linear term gives 10.87572; deleting degree 5 gives an apparent 2.78966 but fails positivity at n=6. Deleting degree 7 also fails there.
Removing the assumed gap exposes negative action near nu=2.21. This remains a bound conditional on that gap and the paper's regularized-crossing framework.
There is no statistical hold-out. The frozen rational witness was checked against the entire allowed spectrum, including modes beyond the search range.
The claimant verifier imports no search module; my certificate uses a different integration method and no claimant modules.

Relative to the paper, the addition is a concrete four-term rational functional, its smaller numerical bound and a complete spectrum certificate within the authors' existing linear-programming approach.
Prior result encountered: Salehi Vaziri's thesis also prints 5.67049. Literature coverage and access limits are recorded in PRIOR_ART.md; no worldwide-record conclusion is asserted.

Caveat / strongest author objection: I reproduced the fully specified one-functional baseline, but the eight pairs and coefficients behind the printed 5.67049 were unavailable in the recovered primary archive. The historical eight-functional calculation itself was not reproduced.

NEXT BAR: recover those eight pairs and coefficients and reproduce 5.67049 with the independent evaluator used here.