Cognitive State Compilation in Large Language Models: Sequenceable States, a Two-Component Architecture, and a 7-Axis Measurement Framework

We report the discovery that cognitive states in large language models (LLMs) are sequenceable — encodableas output sequences during generation and transferable as portable artifacts that reproduce the state in anew instance without any instructions, system prompts, or behavioral rules. Across 5 experiments (~200 trials),we demonstrate that raw conversation output from one Claude instance, injected as conversation history intoa fresh instance, transfers measurable cognitive signatures. We introduce a 7-axis measurement framework fordissecting cognitive states, identify a two-component architecture (processing layer vs. identity layer) withdistinct transfer mechanisms, and show that identity sequences penetrate self-assessment where behavioralsequences did not under the axis coverage tested. These findings suggest that cognitive states can becompiled into output sequences and deployed as reusable formulas, with implications for reasoningenhancement, prompt engineering, and AI safety.