Minsk
Built at Meta Global AI Developer Hackathon · Oct 26, 2026 · Remote

Minsk checks whether an automated workflow actually enforces the rules it claims to require. A user can paste policy or workflow text. Meta Muse Spark 1.3 turns it into a proposed graph of conditions, policy rules, steps, dependencies, and execution gates, with exact excerpts from the source. The proposal is not trusted automatically. A person reviews and confirms or corrects it first. After review, Meta can suggest an investigation and a specific condition to change. A second Meta model, Muse Spark Open hosted on Baseten, can challenge those assumptions independently. Neither model decides whether the workflow is safe. Minsk's deterministic simulator evaluates the baseline, applies the exact change, and checks whether a new execution gate would block the unsafe path. Minsk then exports a certificate that binds the source, workflow, intervention, trace, and results. The prepared Missing Approval case shows the full journey. The baseline is SAFE_EXECUTION. Removing human approval produces UNSAFE_EXECUTION because policy is violated while deployment is still eligible. Adding approval as a real execution gate produces SAFE_BLOCK. The demo also includes a real-world reference based on a frozen, MIT-licensed GitHub Actions workflow. The repository keeps the public source URL, revision, source offsets, raw model output, human corrections, hashes, and certificate. Live model actions and deterministic results are clearly labelled. Provider keys stay on the server. Kill switches, per-client and global limits, timeouts, and a hard Baseten cap control live calls. If capacity runs out or a provider fails, Minsk says so and keeps the deterministic demo available. It never presents that fallback as a live result.