Autopsy Band

Most post-mortems stop at "the model hallucinated" or "the prompt was wrong." Autopsy Band goes deeper — it answers *why* the AI build failed at the architectural decision level, not the symptom level. The system is a 6-agent forensic council running on the Band SDK: Annotator, Verifier, ConfidenceScorer, Reconstructor, ApexSynthesizer, and HumanEscalation. Each agent is a registered Band agent in a shared chat room. Handoffs are Band messages carrying structured metadata — the diagnosis propagates through the pipeline as the council debates, challenges, and synthesizes evidence. The Verifier enforces a hard invariant in code: no claim survives without a verbatim quote present in the original input. If nothing can be validated, the system returns FM-00 — a high-confidence "insufficient evidence" result — rather than hallucinating a diagnosis. Claims that pass the Verifier are scored for confidence and reconstructed into an alternative scenario before the Apex agent synthesizes the final verdict. The council classifies the failure against a 17-mode taxonomy covering spec drift, context collapse, reward hacking, capability overestimation, and more. The web interface is a FastAPI + HTMX server-rendered app. No JavaScript build step. Users paste their failed AI build description and receive a full Autopsy Report in Markdown — with cited evidence, a failure mode classification, and recommended corrective actions. An in-process orchestrator runs the same stage logic as a fallback, so the app works without Band credentials for local development.