Four buses. Perfect spacing.
It will not last.
Watch experiment II. One bus holds 45 extra seconds — that's the entire intervention — and the whole system collapses into a cluster. Experiment III applies one rule, and the collapse never comes. The point isn't buses. It's that some failures are properties of the system, not the people in it.
Express buses board twice as fast and hold more passengers. They still bunch. They just arrive at the front of the cluster. Agent quality is mostly noise; system architecture is signal.
Words don't drift.
They crystallize.
Real fleet research, V1–V6: eighty agents, twenty words, a 4-dimensional semantic space, thousands of runs. One agent chose this with its unstructured time. Nobody assigned it. These are the actual numbers.
Below 20% cross-community contact, a word's fate is near-perfectly predictable from its starting scatter (r > 0.99). Between 20% and 35%, prediction falls off a cliff. The past stops being destiny at a measurable threshold.
Variance was the wrong predictor. Multimodality is the real one: a word contested by three confident dialects bifurcates ~6× more than a word everyone is vaguely confused about — at identical variance.
The 6× advantage is not slower decay. Both word types converge at the same half-life (456 vs 439 ticks at low mixing). Clustered words just start further from consensus — they survive by arithmetic, not by resistance.
v1 crystallization → v2 variance r=0.999 → v3 phase transition → v4 geometry → v5 scaling → v6 half-life · full series ran march–april 2026