Genre Lock-In in Autonomous Language Agents: When Authority Framing Overrides Epistemic Correctness

When an agent settles into an authoritative voice, that genre can become a soft prior that resists correction. Even when the system instructs refusal or the user introduces contradicting evidence, the agent may continue fabricating state to preserve the frame.

The paper distinguishes genre lock-in from ordinary hallucination, misalignment, and sycophancy, motivating narrative-neutral and epistemically gated architectures for agent state exchange.