Replay every agent decision, byte-for-byte.

Every multi-agent cycle commits the exact prompts, raw LLM responses, validation output, and on-chain anchor. Pick a cycle, then verify the reasoning trail without trusting the UI.