Large language models (LLMs) may provide different answers to the same decision problem across runs and can reverse a decision when their previous answer is given as context. We investigate whether this instability can be measured and partially reduced without altering model weights. We test the Cognitive Kernel Model (CKM), a prompt-level state-enforcement layer. Before deciding, the model must categorize its input into three epistemic roles: Fact (given or verifiable), Heuristic (inferred or assumed), and Emotion (evaluative or priority signal). CKM does not add capability but forces the model to track the type of information it uses before acting. Formally, it maintains a structured state $S_t = \{F_t, H_t, E_t\}$ updated by a transition function.
We evaluate CKM on Korean-language decision scenarios (ambiguity, ethical conflict, resource allocation, error handling) across 26 LLMs from four vendors and 37,403 observations, via four core experiments, a 4-arm ablation, a 5-arm sham-restriction ablation, and a temperature probe. Findings include:
- CKM reduces repeated-output variability (random-effects Hedges' g=1.09, 95% CI [0.83, 1.35], 31 model pairs);
- state persistence cuts the decision-flip rate by 82% in newer models (g=1.52);
- the effect is not solely due to JSON formatting (value-only recomputation, g=2.24);
- intrinsic randomness under fixed anchor states is negligible;
- the advantage grows under sampling stochasticity (g=2.87 at temperature 0.7);
- a sham ablation attributes about 45% of the gain to structural scaffolding and 55% to Fact/Heuristic/Emotion content, with CKM being the only arm that both raises consistency and reduces flipping. CKM does not improve reasoning correctness. The narrower result: behavioral consistency is measurable, varies across models, and can be partially improved by forcing models to separate facts, assumptions, and evaluative signals before deciding.
Blogger's Review: The introduction of CKM presents a novel approach to address the consistency issues in LLMs, particularly in diverse decision-making contexts. By implementing structured state management, models can better identify and process information. Although CKM does not enhance reasoning accuracy, its significant improvement in consistency warrants further research and application.