Matrix Globe · Meta Global AI Developer Hackathon

No proof, no bug.

AI agents write code that looks right and is wrong in one detail. Matrix Globe runs a peer review where the reviewer, Meta's Muse Wonder, may only report a defect if it writes a test that fails on the code in an isolated sandbox. Claims it cannot reproduce are thrown away. A builder on Muse Spark revises until every proven test passes, and the run ends in a SHA-256 certificate anyone can check.

1 · Build

Muse Spark writes the code from a plain-language spec.

2 · Prove

Muse Wonder writes one failing pytest per defect. The sandbox runs it: PROVEN, UNPROVEN (discarded) or INVALID.

3 · Certify

Muse Spark revises until all proven tests pass. The canonical JSON of the run is hashed into a certificate.

FinCode-30 results: a review benchmark for finance code

30 finance tasks, 20 held out for testing (rounding, day counts, IBAN, amortisation, FX, XIRR, reconciliation…). Every target already passes its public tests and hides one plausible bug that only hidden tests catch, or is clean code that must not be broken.

Proofs

Task · targetModelsConditionStatusProven / claimsValidCostsha256