Cognitive Governance as a Model-Agnostic Architectural Layer for Large Language Models: A Framework and Preliminary Empirical Evaluation

This paper introduces a cognitive governance architecture for large language models, referred to as the Core. The architecture operates above existing models, without changing model weights, requiring fine-tuning, or depending on a specific vendor.

The paper begins from a practical failure pattern: language models can produce fluent answers while drifting from constraints, accepting user framing too easily, collapsing ambiguity into premature decisions, or losing coherence across extended interactions. These failures are treated here as related expressions of a broader condition called epistemic non-independence: the tendency of a model’s reasoning to follow the nearest pressure signal instead of remaining anchored to evidence, constraints, and the prior governance frame.

The Core is presented as a behavioral governance layer rather than a reasoning improvement or a new model. Its purpose is to discipline the path through which existing model capability is expressed. The architecture is described through three functional pillars, a four-phase reasoning cycle, and a set of behavioral dimensions for contextual meta-reasoning.

The paper reports preliminary evidence from several scenario classes involving constraint handling, strategic ambiguity, behavioral restraint, formal constraint discipline, contextual progression, and crisis-simulation behavior. These results are interpreted as convergent but provisional. They do not establish generalization, and they require independent replication under a unified evaluation protocol.

The contribution of this work is the articulation of a model-agnostic layer in the AI stack: cognitive governance as infrastructure above language models, designed to preserve reasoning discipline under pressure before outputs become operational decisions.