Skip to content

Latest commit

 

History

History
28 lines (25 loc) · 1.91 KB

File metadata and controls

28 lines (25 loc) · 1.91 KB

Claude Science independent audit: re-derivation table

A separate Claude Science session, on separate infrastructure, cloned this public repository from scratch, read the scoring / baseline / operating-point / verifier scripts plus data/gosai_designed/designed_scored.csv, and independently re-derived every committed number. This is the artifact behind the README's "independently reproduced by a separate Claude Science session" line. Every captured row matched to the decimal.

Claim Repo value Science re-derivation Match
Pooled cross-lab AUROC 0.8013 0.8013 yes
Cross-lab AUPRC 0.293 0.2929 (random 0.0629, 4.66x) yes
Base failure rate 6.29% 6.2867% (5,874 / 93,435) yes
Cell-prior-only AUROC 0.678 0.6784 yes
Within-cell macro AUROC 0.747 0.7473 (HepG2 .833 / SKNSH .731 / K562 .678) yes
Joint (cell x method) macro 0.662 0.6624 (14 strata) yes
Per-generator macro 0.726 0.726 (hmc .922, sa .776, sa_rep .695, fsp .625, al .613) yes
GC / length baselines 0.51 / 0.50 0.5113 / 0.5000 yes
Riskiest-2% PPV / enrichment (pooled) 0.49 / 7.8x 0.487 / 7.74x (economics_table.py, positive control against reproduce_flagship.py) yes
Riskiest-2% enrichment, conditioned within (cell x generator) 2.59x macro triage_conditioning_check.py added post-audit
Riskiest-10% capture / lift 43% / 4.3x 42.6% / 4.26x yes
Safest-50% failure rate 1.91% 1.9072% (69.7% reduction) yes

Science's verdict on the flagship: honest, presented more conservatively than it needs to be, no overclaim. In the same audit it independently reproduced the variant-effect extension headline (see docs/variant-effect-extension.md) and caught a reproduction-claim conflation that an earlier self-review had missed. The full account of the Claude Code and Claude Science collaboration is in HOW-BUILT.md.