Synthetic Players

ClaimRegistered · mixed record

P5-2 — persona-direction vs task-text classification

Pooled 45/352 = 0.128 task-consistent; historical verdict preserved; Bayesian proximity to 0.20 is prior-dependent.

Registered conflict-cell pooling · frozen boundary 0.20 · Paper §4.2, §5.1

P5-2 pools registered conflict cells and classifies whether choices follow task text or persona-conditioned direction. The pooled task-consistent share is 90/704 seat decisions across 352 episodes — 45/352 = 0.128 on episode means. The prompt-cluster bootstrap (principal dependence-aware sensitivity) gives [0.071, 0.189].

Prior sensitivity of the Bayesian aggregation

symmetric αposterior median95% intervalP(θ ≤ .20)
0.100.138[0.124, 0.153]1.000
0.250.152[0.135, 0.171]>0.999
0.500.172[0.152, 0.195]0.991
1.000.205[0.182, 0.231]0.329

The α=1 posterior crosses the registered 0.20 boundary, so proximity to the boundary is prior-dependent rather than an independent signal. Every repeated conflict subcell is mixed; only the swap cell is individually persona-dominant — the pooled classification is mechanism-confounded and carried by the word/payoff-confounded swap cell.