Fork the consciousness, or download the project and create your own.

Dushyant Sharma

Gubernaut project (gubernaut.com)

Functionalist What matters is the organisation of the processing, not the material

Dushyant Sharma leads the Gubernaut project, whose 2026 paper examined on this site in the Gubernaut analysis treats the reactive failure modes of language model agents, escalation under provocation, sycophantic drift, perseveration, as failures of propensity that runtime architecture can address. The Gubernaut Cognitive Controller is a Nelson-Narens monitoring and control loop with an unusual design constraint. The object level reads and writes text. The meta level ingests zero tokens, reading only three numbers, intensity, valence, and repetition, and returning a regulating posture. Because the regulator never sees language, there is no injection channel into it by construction.

The validation is as notable as the design. A pre-registered generate-once, judge-many protocol across a four by four matrix of frontier models, each serving as generator and judge, found the regulated arm calmer in thirteen of sixteen cells at p below .05 and fifteen of sixteen by sign, with the effect surviving a lineage-independent judge family. The mechanism is the recovery signature, arousal that integrates under sustained attack and decays, gated by valence, once de-escalation begins, replicated across all four families. The paper claims no consciousness, and its transcripts ship with SHA-256 provenance so the panels can be re-judged.

The position the work defends is that affect-shaped dynamics are an engineering layer, and that the separation between affective regulation and symbolic content can be built thin enough to audit. For consciousness science the result is a boundary marker. Homeostatic valence-like control now exists as validated runtime infrastructure, which sharpens the question the field keeps deferring, what would make such dynamics felt rather than merely functional.

Known for. The Gubernaut Cognitive Controller, zero-token affect regulation, the recovery signature across model families, pre-registered agent evaluation

Coverage on this site

Related researchers