Cognitive Model Organism: The Unboxing of Thinking Process

We establish the Hortus Regis Alti (HRA) Experimental and Observational Protocol for studying a frozen language model as a cognitive model organism. The preparation combines controlled input perturbations, complete native-token preservation, deterministic symbolic probes, contextual process coding, and explicit failure screening. One DeepSeek-R1-Distill-Qwen-7B implementation with INT8 linear layers and BF16 elsewhere produced 176 archived trajectories across four observation configurations of one Bayesian decision problem, containing 2,205,963 generated tokens. Eight baseline pairs regenerated identical overlapping prefixes across 93,016 token positions. Three empirical demonstrations organize the study. Symbolic closure separated from execution termination, including renewed processing after a closure marker. Aggregate consumption separated from item-specific response paths: 36 of 40 four-level item series changed direction. Correct local computation separated from maintenance of task representation: an auxiliary result could be calculated correctly and subsequently rebound as an operand. Complete contextual reading localized 1,926 evidence anchors, including successful repairs and recurrent failures. Measurement checks further showed that a short window could classify a long exact cycle as highly diverse. These findings motivate testable hypotheses of representational control competition, training-register capture, and content-specific effects of symbolic continuation. The contribution is a process-oriented experimental preparation, a reproducible trajectory collection, and a quality-control procedure that connects observable symbolic events to discriminating intervention sites. Precision specifications identify the model strain, execution conditions, measurement units, semantic coding decisions, and regeneration range supporting these claims.