ClaimPrior-sensitive
Between-prompt composition — substantial, dominance prior-dependent
Median between-prompt share 63–71% (Jeffreys α=.5), 47–53% (α=1); plug-in 85–96%.
Variation in the panel is strongly prompt-indexed, but its share depends on the uncertainty model:
| Uncertainty view | Between-prompt share (4 repeated-game cells) | Reading |
|---|---|---|
| Plug-in / conditional bootstrap | 85.5%, 96.1%, 88.8%, 90.2% | conditions on recorded boundary concentration; tends upward |
| Fixed-panel Dirichlet, Jeffreys α=0.5 | medians 63.1%, 70.5%, 66.1%, 66.6% | shrinks six-episode corner cells toward the interior |
| Symmetric α=0.25 | medians 74.8%–82.5% | P(share>½)>0.999 in every cell |
| Symmetric α=1 | medians 47.1%–53.5% | P(share>½) only 0.325–0.700 |
The data support substantial prompt-indexed composition across the sweep, while the stronger “dominant between-prompt share” reading is prior-dependent. The views bracket the claim under opposite conditioning choices; they are not interchangeable estimators.