# Framing A/B

Real decisions. No separate paid startup probes preceded this experiment.

Three shared board seeds per model and framing, one model in seat 1 with three WeightedRandom bots. Each pair uses identical engine seeds; later dice can diverge as choices consume randomness. The 250-turn guard uses actual points to choose a leader. Builds include setup settlements.

Selection rule fixed before paid execution: Keep v2 if it wins more games overall, or if wins tie and it has higher average points without fewer settlement and city builds combined. Otherwise keep v1. This is a small descriptive experiment, not a significance claim.

| Model | Framing | Games | Wins | Average points | Settlements built | Cities built | Fallbacks | Cost USD |
|---|---|---:|---:|---:|---:|---:|---:|---:|
| jev | v1 | 3 | 1 | 5.33 | 7 | 2 | 0 | 0.064425 |
| jev | v2 | 3 | 2 | 9.00 | 15 | 3 | 0 | 0.056973 |
| ms-d1 | v1 | 3 | 0 | 3.00 | 6 | 1 | 0 | 0.039169 |
| ms-d1 | v2 | 3 | 2 | 8.67 | 15 | 3 | 0 | 0.040649 |
| d1 | v1 | 3 | 1 | 5.00 | 9 | 0 | 0 | 0.025649 |
| d1 | v2 | 3 | 2 | 9.33 | 17 | 5 | 0 | 0.021972 |

Aggregate:
- v1: 2/9 wins, 4.44 average points, 22 settlements and 3 cities, US$0.129243.
- v2: 6/9 wins, 9.00 average points, 47 settlements and 11 cities, US$0.119594.

Selected v2. v2 helped under the stated rule.