Plain-language summary.

A running system logs many internal numbers at each step.

This paper turns those numbers into a small set of coordinates, like a map. It then checks

whether the system’s regime labels match regions on that map, and whether regime switches

look like jumps. No meaning is read from generated text; only timing signals (input arrival

and “say” timestamps) are used.

Technical summary.

A tick-resolved window of a real-time “cognitive runtime” execution

(1k-node substrate) is analyzed using principal component analysis (PCA) on internal

numeric telemetry. In the analyzed window (t ∈ [359521, 385094], n = 16,746 ticks), the

first eight PCs explain 89.18% of standardized-feature variance. Regime-change ticks have

higher per-tick displacement in the first eight PCs, ∥∆PC1:8∥2 (mean 6.05 vs. 5.05; Cohen’s

d = 0.47; Mann–Whitney p = 2.9 × 10−157). A microstate Markov model (K=30, built

on the first eight PCs) exhibits multiple slow modes (second eigenvalue 0.9978, implied

timescale τ ≈ 455 steps), consistent with metastable organization. A “content influence”

control uses a lightweight input embedding (hashing + SVD) and finds a measurable increase

in next-step PC predictability (mean ∆R2 = 0.134 for predicting PC1:6(t+1)), while direct

next-step “say” prediction does not improve (AUC change −0.0055). All results are backed

by a self-contained data+code bundle with SHA256 indexing.