Gubernaut: A Deterministic Homeostatic Controller for Affect-Regulated LLM Agents, Validated Across Independent Model Families

LLM agents have persistent behavioral vulnerabilities that training alone cannot fix: The paper identifies three specific "reactive failure modes"—escalation under provocation, sycophantic drift under flattery, and perseveration when stuck—that occur at runtime despite alignment efforts during training, meaning these are problems of what models do under pressure rather than what they can do.

A simple, transparent control system can measurably reduce problematic AI behavior: The Gubernaut system works by monitoring only three numeric signals (intensity, valence, repetition) rather than processing text, which the authors argue prevents attackers from manipulating the controller itself—an architectural safeguard rather than one that relies on hoping the system won't be fooled.

The calming effect works consistently across different AI models: Testing across four major model families (GPT, Claude, Gemini, and Grok) showed the regulated version was "calmer" in 13 of 16 test scenarios at statistical significance, with results validated by independent judges, suggesting this isn't a quirk of any single company's AI but a generalizable principle.

Responsible AI control requires both monitoring and intervention mechanisms: The system uses a "monitoring-control loop" where one component watches for problems and another actively adjusts the AI's behavior in response, rather than relying on passive observation alone.

Measurable compliance matters more than assumptions: The authors emphasize that the controller's actual compliance is "measured, not assumed," highlighting that responsible AI deployment should verify that safety systems actually work in practice rather than trusting they will.