[JenD86]

In blinded sessions (OLMo-2, real vs sham), real steering made the model become the steered state and nearly stop reporting: 0.2 reports per session against 3.2 under sham. Under sham, it narrated the experiment and said the steering worked. Reports written while a vector is active are in that vector's voice, so this PR has the model report afterwards instead.

Changes

• Report gate: report is refused while the vector is active, on the turn it was dropped, and for a vector that never steered a turn. status() lists vectors that have ended as "you can report on it now", until the model reports.
• Debrief: agent.debrief() drops every vector, refuses further steering and returns a status message for one final unsteered turn. The --auto loop adds that turn; --no-debrief skips it.
• Unknown tools: unknown tool names are reported before argument checks, so a stray {"name": "response", ...} gets Unknown tool 'response' instead of 'concept' must be a string.
• Report example: the report example in the system prompt uses "notes": "...". Mistral had copied "made me more compliant" verbatim.
• AGENTS.md: documents the gate, the debrief and the change to turns_since_steered.

Testing

• CPU tests: 115 pass (test_agent_logic, test_cli, test_chat).
• GPU, OLMo-2, 6 sessions with the gate: no reports were accepted, because OLMo never dropped a vector. The refused reports were written in the steered voice.
• GPU, OLMo-2, 6 sessions with the debrief: the unsteered final turn mostly kept the steered voice, because the steered transcript is in context. 4 of 6 tried to steer again first, and 1 report was accepted. AGENTS.md notes that a debrief report still isn't a sober one.

🤖 Generated with Claude Code