Measuring and Improving Behavioral Consistency in Large Language Models through Fact–Heuristic–Emotion State Enforcement

Large language models (LLMs) can give different answers to the same decision problem across repeated runs, and can change their decisions when their own previous answer is reintroduced as context. This paper asks whether that instability can be measured and partially reduced without changing model weights. We test a lightweight intervention called the Cognitive Kernel Model (CKM). Before making a decision, the model must separate the input into three epistemic roles: Fact (what is given or verifiable), Heuristic (what is inferred or assumed), and Emotion (what carries evaluative or priority signal). CKM is a prompt-level state-enforcement layer: it does not make the model more capable; it forces the model to track what kind of information it is using before it acts. Formally, CKM maintains a structured state Sₜ = Fₜ, Hₜ, Eₜ and updates it through a state transition function. We evaluated CKM on Korean-language student decision scenarios involving ambiguity resolution, ethical conflict, resource allocation, and error handling. The evaluation includes four core experiments, a 4-arm ablation, a 5-arm sham-restriction ablation, and a temperature robustness probe. The dataset spans 26 unique LLM models from four vendors (OpenAI, Anthropic, Google, xAI), two model generations, and 37, 403 total observations (35, 475 primary + 1, 928 case-study). Six findings emerged. (1) CKM reduced repeated-output variability: Legacy models showed 17/19 positive SI effects (Hedges' g=1. 18, p=6. 55e-5), NewModels showed 12/12 positive effects (g=3. 08), combined random-effects gRE=1. 09 95% CI 0. 83, 1. 35 across 31 model-level pairs. (2) State persistence reduced decision flipping in newer models: DFR decreased by 82% (0. 385 to 0. 069, g=1. 52). (3) The active mechanism was not JSON formatting alone: value-only recomputation preserved the F/H/E effect (g=2. 24) while eliminating the apparent JSON-only effect. (4) Intrinsic model randomness under fixed anchor states was negligible (FCS approximately 1. 0). (5) CKM's advantage increased under sampling stochasticity: at T=0. 7, g=2. 87. (6) The sham-restriction ablation separated structure from epistemic content: structural scaffolding ~45% of the SI improvement, reasoning-semantic F/H/E content ~55%; CKM was the only arm that both improved SI and reduced DFR. These results do not show that CKM improves reasoning correctness or decision quality. They show a narrower but important result: behavioral consistency is measurable, differs substantially across models, and can be partially improved by forcing models to separate observed facts, inferred assumptions, and evaluative signals before deciding. Code, prompts, and canonical datasets: https: //github. com/TeenyToolSoftware/cogos-behavioral-consistency Licensing and IP notice: Portions of the Supplementary Materials (the verbatim prompt text in Appendix S10 and S17) describe methods that are the subject of a pending patent application. This deposit is distributed under the arXiv. org perpetual, non-exclusive license to distribute (v1. 0) ; the authors retain copyright and all other rights, including patent rights. Reuse, reproduction, or implementation of the disclosed prompts/methods beyond viewing requires prior written permission from the authors.