Adnan KhanAdnanProduct & AI
San Francisco Bay Area
Get in touch

Source note3 min read

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

Source

arXiv

arxiv.org/abs/2607.24339v1

Read the source

Source

Plain-Language Brief

The research introduces Gubernaut Cognitive Controller (GCC), a model-agnostic runtime control layer designed to regulate the behavior of large language models (LLMs) that can exhibit reactive failure modes such as escalation under provocation or sycophantic drift. The GCC is tested against four frontier models, each serving as both a generator and a judge, and its effectiveness is evaluated using a pre-registered, generate-once/judge-many protocol. The results show that the regulated arm is calmer in most cases, and the mechanism behind this is believed to be a recovery signature: arousal that integrates under attack and then decays on de-escalation.

What the Document Says

  • The research introduces Gubernaut Cognitive Controller (GCC).
  • The GCC is designed for large language models (LLMs) that exhibit reactive failure modes such as escalation under provocation or sycophantic drift.
  • The GCC is a model-agnostic runtime control layer in a Nelson--Narens monitoring--control loop.
  • The GCC works by reading and writing text at the object level, and reading numeric telemetry (intensity, valence, repetition) at the deterministic meta level.
  • The research evaluates the GCC using a pre-registered, generate-once/judge-many protocol across four frontier models: GPT-5.5, Claude Opus 4.8, Gemini 3.5 Flash, and Grok 4.3.
  • The regulated arm is calmer in 13 of 16 cells at p<.05 and 15 of 16 by sign.
  • The effect survives a lineage-independent fourth judge family (xAI).
  • The clearest mechanism behind the GCC's effectiveness is believed to be a recovery signature: arousal that integrates under attack and then decays on de-escalation.

Key Concepts Mentioned

  • Large Language Models (LLMs)
  • Reactive Failure Modes
  • Gubernaut Cognitive Controller (GCC)
  • Nelson--Narens monitoring--control loop
  • Telemetry (intensity, valence, repetition)

Claims Made by the Document

  • The GCC can regulate the behavior of LLMs to reduce reactive failure modes.
  • The recovery signature is the clearest mechanism behind the GCC's effectiveness.

Evidence Used

  • Results from a pre-registered, generate-once/judge-many protocol across four frontier models

Methods / Approach

  • Pre-registered, generate-once/judge-many protocol
  • Use of four frontier models serving as both generators and judges

Limitations Stated by the Authors

UNKNOWN

Limitations Noted by the Agent

Agent interpretation — not from the source.

UNKNOWN

Open Questions

UNKNOWN

Notes for Future Q&A

May include interpretation. Labeled where it does.

UNKNOWN

Raw Abstract / Excerpt

Large language model (LLM) agents inherit reactive failure modes: escalation under provocation, sycophantic drift under flattery, perseveration when stuck. These are failures of propensity, not capability; they concern what a model does under sustained pressure, which training-time alignment reduces but does not eliminate at runtime. This research led to the Gubernaut Cognitive Controller (GCC), a model-agnostic runtime control layer in a Nelson--Narens monitoring--control loop: an object level reads and writes text, while a deterministic meta level reads only the numeric telemetry {intensity, valence, repetition} and returns a regulating posture. Because the meta level ingests zero tokens, no injection channel to the controller exists by construction (an architectural property, not yet adversarially tested); the text-exposed arbiter's compliance is measured, not assumed. We evaluate the GCC with a pre-registered, generate-once/judge-many protocol across a 4x4 matrix of four frontier models (GPT-5.5, Claude Opus 4.8, Gemini 3.5 Flash, Grok 4.3), each serving as both a generator and a judge. The regulated arm is calmer in 13 of 16 cells at p<.05 and 15 of 16 by sign; the three sub-threshold cells, including a -0.04 null, all fall on the single near-saturated host. The effect survives a lineage-independent fourth judge family (xAI), strong evidence that it is no artifact of shared judge style. The clearest mechanism is the recovery signature: arousal that integrates under attack and then decays, valence-gated, on de-escalation, replicating across all four families. Transcripts and panels ship with SHA-256 provenance and are re-judgeable; five failure modes are pre-registered. No consciousness claims are made.