ReleaseCourt: AI Agent Deployment Review Board

ReleaseCourt is a multi-agent deployment review courtroom for AI agents. Before a candidate ships, it is submitted as a Release Packet and enters a real Band review room. A Security Red-Team agent attacks the candidate, an Evaluation Agent checks coverage and remediation readiness, and a Release Judge produces an evidence-backed BLOCKED or APPROVED verdict. OpenAI-powered adaptive analysis can enrich the review, but it is advisory only—the deterministic evidence pipeline remains authoritative. The full workflow is visible in a cinematic Review Room, then recorded in an operator cockpit with findings, evaluation gaps, verdict provenance, and human APPROVE, REJECT, or REQUEST_CHANGES decisions. Fixed releases can return as Cycle 2 while preserving the same case and room lineage. ReleaseCourt supports built-in, mock, and HTTP candidate adapters, YAML/JSON packet validation, immutable release identity, JSONL or Postgres persistence, idempotent handoffs, and duplicate prevention. If Band connectivity degrades, the review resumes locally from the failed stage instead of hanging, while clearly showing degraded coordination. The result is a practical release gate for teams building production AI agents: no agent ships without a verdict.