What problem does it solve? Self-improving agent pipelines accumulate guardrails and lessons, but without structured retrospective analysis there is no way to know whether a guardrail actually reduced its target failure, whether a rule was ever baked into a playbook, or when a recurring failure class deserves escalation to a deterministic check. ## Core Features & Use Cases - Guardrail Effectiveness Computation: Compares failure event rates before and after a guardrail's activation date using per-issue JSONL event logs and state-machine metrics, classifying results as Confirmed, Partial, Pending, or Ineffective. - Cross-Spec Pattern Detection: Groups failures across multiple specs by class to surface unguarded recurring patterns and prioritize new guardrail proposals. - Grounded Verification & ACE Curation: Verifies guardrail rules exist verbatim in agent playbooks, then applies an archive/escalate/strengthen decision tree following the ACE Generation-Reflection-Curation lifecycle. - Use Case: After an audit verdict, the Self-Improver agent runs these recipes to persist a new guardrail record in references.md, verify prior guardrails still work, and escalate any rule that was ignored despite being documented. ## Quick Start Ask the Self-Improver agent to run the retro-analysis recipes against the current pipeline event logs and report guardrail effectiveness with any unguarded failure patterns.
