Gubernaut: Runtime Controller for Affect-Regulated LLM Agents

A team of researchers has introduced the Gubernaut Cognitive Controller (GCC), which serves as a model-agnostic runtime control layer for large language model (LLM) agents. This innovative GCC tackles reactive failure modes, including escalation under provocation, sycophantic drift due to flattery, and perseveration when encountering obstacles. These issues stem from propensity rather than capability, and while training-time alignment helps, it does not fully resolve them during runtime. The GCC functions within a Nelson–Narens monitoring–control loop, where the object level processes text and the deterministic meta level analyzes numeric telemetry (intensity, valence, repetition) to establish a regulating posture. Notably, the meta level does not process tokens, eliminating any potential injection channels to the controller, a feature that has yet to undergo adversarial testing. Compliance of the text-exposed arbiter is assessed rather than presumed. The evaluation of the GCC utilized a pre-registered, generate-once/judge-many approach across various independent model families, as outlined in arXiv paper 2607.24339.