Loops HouseFixseal

NEAR HACKATHON
Team
1 member- MOOwner
Moriki Kamio
Overview
AI agents now write the fixes for code that holds money, faster than people can review them. Tests are not enough: in our runs, AI fixes passed every attack and regression test while their proofs were still wrong.
Fixseal is the gate an AI fix must pass before it can change how an agent spends. The judge reruns the original exploit and variants, reruns regression tests, rejects policy tricks (sorry, extra axioms, edited specs, weakened tests, hard-coded inputs) and has the Lean 4 kernel check the AI's proof against a human-approved spec. Anyone can rerun the judge from the bundle.
NEAR makes the verdict binding. A NEAR contract holds the agent's spending rule; the agent's function-call key can only call guard_spend; the rule can be replaced only by the exact expression an accept record covers. On testnet: an overflow attack pays past a 1000 yoctoNEAR cap; replacement with no record or a reject record fails; the accepted fix goes in; the same attack is refused; normal spending works.
Live, in the demo video: Codex fixes a blind copy (answer removed) while we record, is accepted on the first try (attacks 12/12 → 0/12, regressions 16/16, Lean proof checked), and that exact rule is installed on testnet. Earlier recorded runs: Devin and Codex passed every test yet got PROOF_REJECTED first, then reached accept; DeepSeek V4.1 Flash on NEAR AI Cloud also produced an accepted fix (rejudged by us).
Check it in a browser, no wallet or keys: morikikamio.github.io/fixseal reads the live contract, matches the code hash to the CI build, checks the rule came from an accept record and the key is limited to guard_spend, and runs the stored rule on any amount you type.
Limits: the contract trusts the record submitter; the Lean-to-contract match is tested on 1,028 inputs, not proved; attestation is mock. Next: more spec types, independent verifier signatures, TEE re-execution.
Try: scripts/judge-check.sh --stage a (no install).