Last released Oct 8, 2026
Test a chat agent through Claude Code or Codex without the model retyping the conversation, and audit that it held.
Deterministic admission for LLM-extracted facts: admit, flag for verification, or reject, with receipts.