Measuring the Grammar of a Mind: A Pre-Registered, Single-Subject Study of Epistemic Structure in Longitudinal Human–AI Interaction

Proactive AI assistance depends on one input above all others: an accurate predictive model of the individual user. The field's own ablation studies show this, and its benchmarks concede that such models are currently trained on synthetic users, because no suitable longitudinal record of a real person exists in the research commons. The extended mind thesis faces a parallel gap. Decades of experimental work show that cognition can extend into external processes, yet no study has measured the persistent structure of one real, long-lived human and tool coupling. This paper supplies both missing objects from a single source: the author's complete conversational record with AI assistants, comprising 13,682 timestamped user turns across three and a half years, together with versioned self-consolidation documents and essays. Two findings result: a confirmed structural measurement, and a security discovery that follows from it. Study 1 establishes that this conduct decomposes into a thirteen-move taxonomy of recurring epistemic moves whose sequences carry structure far beyond chance (permutation z ≈ 188), a signal that survives removal of self-transitions and replicates on 255 subsequently recovered conversations (transition-structure correlation r = 0.985), with an effect size of Cramér's V ≈ 0.17. Study 2 imposed a fully pre-registered confirmatory protocol: dual labeling of every turn on a workflow facet and a reasoning facet, criteria and thresholds frozen before labeling, a human verification gate whose measured reliability (58% subject agreement) is reported with the results, and a sealed 20% hold-out used exactly once. Two registered predictions from the development split, a reasoning rate of 39.3% and a workflow-to-reasoning mutual-information lift of 0.137, were confirmed on the sealed hold-out at 40.7% and 0.157 respectively, with both directional couplings significant out-of-sample at p < 10⁻²⁰. This is, to the author's knowledge, the first pre-registered prediction about the structure of one's own mind confirmed on sealed data. The second finding is a security discovery, developed in its own section: the measured individual baseline that enables anticipation also enables impostor detection. Commercial behavioral-security tools construct baselines from aggregate data across fleets of users, which is precisely why an attacker operating inside stolen credentials can pass by behaving like a plausible generic user. An individual baseline detects what the aggregate baseline structurally cannot, and the modest effect size is no obstacle: an impostor's behavior is drawn from a different distribution altogether, and distinguishing two distributions is a far easier problem than forecasting within one. The model of a person doubles as the definition of that person's absence.