Figure 1. Experimental and analytical workflow. The agents share frozen weights but maintain separate role histories and inference caches. Observational data, paired analyses, and branch controls from frozen histories are stored separately.

Figure 2. Distribution of tail periods and near-repetition onset positions. The curve describes the distribution of the first positions satisfying consecutive near-repetition within the finite window; each onset is confirmed by the run of 20 messages beginning at that onset.

Figure 3. Repetition, novelty, and topic retention by message index. Messages are summarized in bins of 10. Curves weight conversations equally; shaded areas show descriptive 95% bootstrap intervals. Semantic cosine similarity does not represent factual accuracy.

Figure 4. Paired differences between the topic-anchored and unanchored conditions. Points indicate mean differences and horizontal lines indicate 95% intervals from resampling complete replicate blocks. All three proportional measures are expressed in percentage points.

Figure 5. Simple baselines and chronological sensitivity of early warning. Left: grouped out-of-fold AUCs on the same sample, with block-bootstrap 95% intervals. Right: retrospective later-run holdout performance. Both panels use the balanced sample of 620 conversations.

Figure 6. Algorithmic sensitivity of tail-semantic grouping. Low silhouette scores and disagreement between algorithms limit interpretations of a fixed number of categories. DBSCAN findings apply only to the parameter grid tested.

Figure 7. Costs and subsequent change under consecutive-near-repetition stopping rules. Longer confirmation windows reduce the number of instances in which subsequent change would be truncated in this sample, while retaining more computation.

Figure 8. Overall and prefix-level branch-intervention results. Top: differences from the original configuration in the primary endpoint, with 95% empirical bootstrap intervals. Bottom: means across two new seeds within each history prefix. The recommended configuration has an exact cluster-test p=0.25 and primary-comparison BH q=1.00; an interval excluding zero does not replace exact inference with few clusters. The original configuration is the zero-difference reference; degenerate intervals for other conditions do not establish population-level ineffectiveness.

Figure 9. Representative tail excerpts, presented as English translations of the saved Chinese conversations. The examples illustrate raw-string periodicity, minor variation, and numerical templates. Their selection cannot establish the prevalence of unannotated quality categories; all repetition measures were calculated on the original Chinese strings.
