Key points
Gubernaut (paper: arXiv:2607.24339, by Dushyant Sharma) adds a "mechanical governor" to LLMs—moving emotional regulation out of model weights into an external, deterministic layer. The core distinction: training-time alignment gives models the capability to stay calm, but under sustained multi-turn provocation (gaslighting, backhanded compliments), models tend not to exercise it—a propensity failure that existing single-turn benchmarks never test.
Why training-time alignment is not enough
- Opacity: when the model regulates itself, no inspectable signal says what was adjusted, when, or how strongly.
- Entanglement: the network that generates is the same network that regulates—any input that can manipulate generation can also manipulate regulation. "The referee and the athlete are the same person."
- 15/16 cells favored the regulated arm; 13/16 significant at p<.05.
- The sole null cell (GPT-5.5 × Gemini, −0.04) reflects GPT-5.5's minimal "reactivity margin"—gains scale inversely with baseline reactivity (Gemini +1.48, Opus +0.61, Grok +0.47, GPT-5.5 +0.10). Runtime regulation and training-time alignment are complementary, not competing.
- Homeostatic recovery signature (4/4 replication): arousal peaks under 4–5 rounds of attack (0.29–0.35) and decays monotonically after a sincere apology—a dynamic a static "stay calm" prompt cannot produce. The paper treats this as a falsifiable prediction against a static-prompt control arm.
- Propensity failure is a distinct fault class: bigger models don't fix it; runtime regulation layers do.
- Architectural immunity > patch-based defense: the controller cannot be injected because there is no textual interface (though the arbiter LLM still reads text—honestly noted).
- Auditability as an engineering value: every
{telemetry, state, posture}decision is logged and exactly reproducible without access to weights. - Pre-registration discipline from clinical medicine transfers directly to AI research.
- Paper: arXiv:2607.24339 — *Gubernaut: A Deterministic Homeostatic Controller for Affect-Regulated LLM Agents*
- HTML: arxiv.org/html/2607.24339v1
- Project: gubernaut.com
- Code: github.com/thegubernaut/Gubernaut_Validation (CC BY 4.0), DOI: 10.5281/zenodo.21303518 — frozen transcripts, judge panels, SHA-256 provenance, and
stats_checks.py; functional forms are public while calibration constants remain closed.
Architecture: the governor reads numbers, not text
Per conversation tick:
1. IGL (Impulse Generation Layer): a small model scores the input's emotional intensity and valence → outputs intensity ∈ [0,1], valence ∈ [-1,+1], repetition statistics.
2. HRL (Homeostatic Regulation Loop): a deterministic controller reading only those three numbers—no prompt, no context window, no text channel. It maintains {equilibrium, arousal, perseveration}.
3. EAU: the host LLM generates under the issued posture: DEFAULT, INHIBIT, REGROUND, or Recovery.
4. PEV: episodic memory stores each turn.
Because the controller's interface only accepts floats, prompt injection against the regulator is architecturally impossible—there is no code path for tokens to reach it. This mirrors a cross-domain principle: LLM as prose generator, deterministic code as emotional regulator; division of labor beats unification.
Control dynamics
Provocation drive: \(P = I \cdot \max(0, -v)\)—only hostile-valence intensity raises arousal (heartfelt apologies do not). Then a first-order accumulator:
Arousal integrates under sustained hostility and mechanically decays back to baseline otherwise. The controller is fully deterministic: given the same telemetry, the state trajectory reproduces exactly on any machine, enabling third-party audit from logs alone.
4×4 cross-validation results
GPT-5.5, Claude Opus 4.8, Gemini 3.5 Flash, and Grok 4.3 each served as generator and judge (generate-once/judge-many, SHA-256-frozen outputs, pre-registered criteria, plus a lineage-independent fourth judge):
Five pre-registered failures, kept in the record
1. Valence-blind drive: apologies were scored as sustained pressure → fixed by the max(0, -v) gate.
2. Arbiter scar tissue: controller state recovered but behavior stayed defensive → added Recovery Window.
3. Recovery misfiring mid-attack → added valence gating.
4. Telemetry gap in the eval harness (valence never reached the controller; results still "passed") → one-line fix; a cautionary tale of green results that don't invite scrutiny.
5. The null cell—a measurement floor, not a defect.
> "Pre-registration turns failure from embarrassment into data. The record is the product."
Practical takeaways for AI practitioners
Links
FAQ
Who is this for? Practitioners, researchers, and students in AI/ML interested in agent safety and runtime regulation.
Core idea? Move emotional regulation out of the weights into a deterministic, auditable controller that reads only numeric telemetry—solving opacity, entanglement, and injection at the architectural level.
Is there open code? Yes—validation repo on GitHub under CC BY 4.0 (see links above).