Open-face Chinese poker
Pineapple OFC
Two seats, five rounds: place each dealt card into back, middle or front, or discard it, against a house seat that is first-legal, random, or the search itself; a board fouls unless back ≥ middle ≥ front, and a fouled seat loses every row. The deal is the seed, so every policy sees the same cards and never the house's hand.
| policy | kind | n | win | return | points/100 | fouls | scoops | royalty | Brier foul · win | house | plies | off-list | cost | record |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| mcts:256@arena/1 | house | 30 | 100% | 1.024 [1.01, 1.04] | +686.7 | 0% | 93% | 1.20 | — | first | 17.0 | 0 | $0.0000 | 91db980caec2… core db1b2b09 |
| mcts:256@arena/1 | house | 30 | 90% | 0.819 [0.56, 1.02] | +526.7 | 0% | 73% | 1.10 | — | random | 17.0 | 0 | $0.0000 | 01b113f56f60… core db1b2b09 |
| first_legal@arena/1 | house | 30 | 40% | 0.267 [0.03, 0.50] | +193.3 | 53% | 40% | 0.00 | — | first | 17.0 | 0 | $0.0000 | 481396d5c4eb… core db1b2b09 |
| random@arena/1 | house | 30 | 23% | 0.067 [-0.13, 0.30] | +23.3 | 77% | 20% | 0.00 | — | first | 17.0 | 0 | $0.0000 | f3bee2f9eb3f… core db1b2b09 |
| deepseek/deepseek-v4-flash@arena/1 | llm | 10 | 10% | 0.100 [0.00, 0.30] | +60.0 | 90% | 10% | 0.00 | 0.639 · 0.279 | first | 17.0 | 21 | $0.0231 | a51cbf7d31b7… core db1b2b09 |
Edition 2026-10, 2026-10-16. Scaffold arena/1. Seeds 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30. Spend on this edition $0.02. Every row is a signed env-scorecard; its link re-checks the seal. off-list counts a model's answers that named no legal action and fell back to the first legal one. The bracket after a return is a 95% bootstrap interval over the seeds. points/100 is OFC points per hundred hands (rows ±1, a scoop ±3, royalties; a foul loses every row); Brier scores the beliefs a model stated at each placement (its probability of fouling, of winning) against what happened: 0 is perfect, 0.25 is a coin; house names the seat it played against.
The episode is the tape
A sealed episode carries its seed, its level, its actions and a state digest per ply. Anyone with the core at that hash replays it and compares every ply: the first digest that differs is the divergence; none is bit-exact. That is how a bug found by a model comes back as something you can reproduce.
curl -s https://api.s2ar.dev/v1/env/tape/80a861bc822df4a460bfce24122d096ea9302590fbf78937fb61a29cab49232b python3 tools/arena.py replay 80a861bc822df4a460bfce24122d096ea9302590fbf78937fb61a29cab49232b