ProofLoop_enhanced

ProofLoop helps developers building visual agents check whether a physical setup satisfies a reviewed requirement. It turns requirements into inspectable checks, evaluates source-bound observations as PASS, FAIL or UNKNOWN, requests missing evidence, and proposes a bounded correction when the observed state supports one. Proposed actions never become measured evidence. The event-built system combines Meta models with a deterministic checker, freshness and stability rules, durable inspection jobs, bounded follow-ups, recovery receipts, request accounting, a read-only MCP interface and a React workbench. The 2:50 demo shows live Spark task drafting and a separate 16-call SAM run on owned prerecorded footage under an already approved cutlery contract. That run produces PASS -> FAIL -> UNKNOWN -> PASS. Judges can also explore a key-free public replay across 32 reviewed checkpoints. This is a bounded placement-verification prototype. The recorded movements predate the planner proposal; an operator-followed correction, robot execution and production reliability remain unvalidated.