Gruya

CodeProof is a static analyzer for TypeScript/JavaScript repositories that publishes only what its evidence gate can defend: every citation is machine-checked against the analyzed commit, a second Meta model tries to refute each claim, and stronger findings earn a generated test (and, where possible, a minimal patch) that flips the result. Rejected and uncertain candidates stay visible in rejected.md. On the frozen v0.2.3-final run — pre-registered, budget-capped, hand-checked and verified by two independent external models — CodeProof found 7/8 calibration defects in both single runs (8/8 dual) with 100% verified precision, 9/16 injected defects on real code (chunked), and 12 hand-confirmed defects across 5 new OSS repositories, at $15.7163 against a $40 cap. 13 findings are filed upstream, all open (one already has a community fix PR). Every number is in reports/final/. After the freeze, a separate exploratory round repeated the measurements (106 paid attempts, $50.10): the weakest published number reproduced (historical recall 2/9 -> 6/30 = 20.0%), and the repeats showed where the gate is not perfect (1 verified false positive in 7 single runs; 1 trap leak in 3 chunked runs) — published as measured, not hidden.