Why
The current replay validation path is stronger than pure LLM-judge validation, but it still replays against installed SKILL.md surfaces and routing phrases instead of the real host/runtime behavior.
PR #96 established the replay provenance plumbing and fixture-backed groundwork. The next step is replacing that simulated replay with actual host/runtime replay where feasible.
Scope
- implement real host/runtime replay for routing validation
- preserve
structural_guard and llm_judge as explicit fallback modes
- keep validation provenance explicit in audit, SQLite, and UI surfaces
- document the host replay flow in
skill/Workflows/Evolve.md and ARCHITECTURE.md
Done when
- routing validation can run against real host/runtime behavior instead of surface-only simulation
- fallback behavior is explicit and tested
- audit/UI provenance still distinguishes replay vs judge vs structural-only outcomes
- regression tests cover replay success, fallback, and unavailable-host paths
Why
The current replay validation path is stronger than pure LLM-judge validation, but it still replays against installed
SKILL.mdsurfaces and routing phrases instead of the real host/runtime behavior.PR #96 established the replay provenance plumbing and fixture-backed groundwork. The next step is replacing that simulated replay with actual host/runtime replay where feasible.
Scope
structural_guardandllm_judgeas explicit fallback modesskill/Workflows/Evolve.mdandARCHITECTURE.mdDone when