"""Agentic Launch Readiness & Misclassification Triage Subsystem. Motivating failure: a pedestrian misclassified as a construction cone at an observed ~0.01% rate. Before treating that number as real, the system must establish denominator, sample size, confidence interval, statistical significance, safety criticality and novelty. Core architectural rule (enforced structurally, see policy.py and agents/base.py): AI agents discover, analyze, summarize and RECOMMEND; deterministic services own metrics, statistics, policy enforcement, launch gates and audit. No LLM output ever directly authorizes a launch decision — final authorization is deterministic policy + recorded human approval. NOTE ON PROVENANCE: the referenced workflow image was not provided to the implementing agent. The five-layer methodology (FAILURE DETECTION -> EVIDENCE AGGREGATION -> FAILURE ANALYSIS -> LAUNCH DECISION -> LEARNING FLYWHEEL) was reconstructed from the written specification and is marked as such. Module map: models.py shared pydantic schemas (FailureEvent, stages, enums) store.py JSON store under runs/agentic/ + append-only audit log data.py deterministic synthetic populations (scenes + rate log) evidence.py Failure Evidence Graph (typed nodes, evidence statuses) snippets.py reproducible failure snippet packages concentration.py uniform-vs-concentrated distribution analysis policy.py deterministic, hash-versioned stop-ship policy engine scorecard.py AgenticSafetyScorecard (leadership retrospective) review.py human-review decisions, mandatory triggers, governance flywheel.py evaluation-suite creation + contamination guard pipeline.py staged five-layer orchestrator (explicit state) worked_example.py deterministic pedestrian->cone walkthrough agents/ the eight agents (advisory only, LLM optional) api.py REST surface under /api/agentic """ METHODOLOGY_PROVENANCE = ( "Reconstructed from the written specification; the referenced workflow " "image was not provided to the implementing agent." )