KEYSTONE

Every AI coding tool says "I fixed your bug." KEYSTONE hands you a signed record that proves it, and you can re-check that record on your own machine without trusting us.

#1 of the CAIN-42 Top 25 live fixing: off ledger: verifies

The idea in one paragraph

The frontier AI model is a commodity. What is scarce is proof. KEYSTONE closes the loop around the model: a deterministic detector, not the model, decides there is a defect; the model writes a fix and a test; an independent reference test the model never saw must fail on the original file and pass on the fixed one, inside a sandbox with no network; the detector must re-scan the fixed file clean; and only then is the result written as VERIFIED_FIX, hashed, chained to the previous record and signed with Ed25519. Anything short of that is written as NOT_VERIFIED, with the reasons, and is never presented as a fix.

The five stages

1. Detect. FALSEWORK (deterministic AST and regex rules, no model, no network) names the defect, the line and the grade.
2. Act. An AI model (Claude Opus 5.5 in every recorded run) returns the fixed file and its own test as structured output. The code is passed to the model as data, not instructions.
3. Prove. In a bubblewrap sandbox (no network, read-only system, 256 MB, 20 s) the model's test and the independent reference test run on the original and on the fix. FALSEWORK re-scans the fix.
4. Sign. The outcome becomes a canonical-JSON record: SHA-256 hash, linked to the previous record's hash, signed with Ed25519.
5. Export. The ledger downloads as a zip with a standalone verifier that refuses to say VALID unless you pin the key yourself.

What you can do on this site

FIRME AI CORPORATION · CAIN-42 is a Byzantine trust fabric. Numbers marked "our own check" were run by us, not by an independent party. Nothing on these pages is a certification.