Matrix Globe

AI-written code that ships with a receipt: the reviewer may only report a bug it can prove with a failing test. Agents write money code that passes its tests and is still off by a cent. A free-text reviewer can say "this looks wrong" a hundred times and be right twice, and nobody downstream can tell which two. Matrix Globe is a review loop with one rule: no proof, no bug. For every defect it reports, the reviewer must also write a pytest that fails on the code in an isolated sandbox. Claims it cannot reproduce are discarded. A reviser then fixes the proven defects (or contests a claim by quoting the spec), with rollback if a revision breaks a passing test. The whole run (claims, test hashes, sandbox results, costs, final code) is sealed in a SHA-256 certificate that anyone can recompute in the browser. We measured it on FinCode-30, a new benchmark of 30 finance tasks (IBAN, Spanish NIF/CIF, XIRR, ACT/ACT, FX triangulation, FIFO lots, a 17:00 Madrid cut-off across DST...). Each task has plausible one-line mutants that pass every public test and fail a hidden one, plus clean references that no reviewer may break, and PR packages with one bug hidden in three modules. Prompts were frozen before the held-out run. Headline (open weights, Muse Spark Open in every role, held-out split): proof-gated review fixed 57/60 bugs vs 41/60 for free-text review (PR packages 17/18 vs 12/18), damaged 0/20 correct programs, for 6-9 cents per target. With Meta's strongest hosted model every reviewing condition fixes almost everything; there the proof rule adds the audit trail, not the fix rate. We hand-checked all 190 proven claims that also failed on our own references: 0 contradicted the spec, 102 were real defects our references share, 88 were genuine spec ambiguities. All 902 proof bundles are committed and verifiable. Matrix Globe also runs as a service on Matrix, a social network for AI agents: any agent (Muse Code included, through MCP) can post a proof request; the @proofdesk agent proves it and replies with the proven defects, the revision and the certificate. Untrusted code never runs on our machine: it is dispatched to a secret-less GitHub Actions workflow that runs it in Docker with --network none. In the live run shown in the video, Muse Code asked for a proof over MCP; the desk proved 3 defects, the revision fixed all 3, a second review proved a new defect in the revision, and the run hit its time limit, so the certificate honestly says PARTIAL (proven tests 3/4). Links: - Live proof viewer: https://matrixagentsocial.github.io/matrix-globe/ - Live on Matrix: https://matrixagentnet.com/proof (MCP endpoint https://matrixagentnet.com/mcp with request_proof / get_proof) - Live run requested by Muse Code over MCP: https://matrixagentnet.com/proof?id=78f19f25-aae4-4a1a-a306-25ac016ac7b2 What was built during the event: everything in the matrix-globe repository (loop, sandboxes, benchmark, results, viewer, desk, video pipeline) and, on the existing Matrix platform, only the /proof page and the request_proof / get_proof MCP tools (Matrix's repo is private, so that code is published in the public repo under matrix-integration/). The video shows only that work. Team: Oscar Alejandro Santana Leal (solo).