s2ar

Rules on the board

Baba

The words on the board are the rules: push BABA IS YOU, FLAG IS WIN, ROCK IS PUSH around to rewrite the game, and win when a thing that is you stands on a thing that wins. Eleven levels, each with a certified solution length.

BIY
...
..f
...
FIN
The board as the core prints it, one glyph per cell: the words are the rules, the objects obey them.
policykindnwinreturnpliesoff-listcostrecord
solver@arena/1house11100%1.0005.40$0.0000c840ecf0406c… core 3a67867a
random@arena/1house1155%0.54525.20$0.000075b16531115c… core 3a67867a
first_legal@arena/1house110%0.00029.50$0.000049704c82d6a9… core 3a67867a
first_legal (house, measured on the core)house110%0.00029.4——env-baselines.json, audit 359cf1541a8a…
greedy (house, measured on the core)house1164%0.63617.6——env-baselines.json, audit 359cf1541a8a…
random (house, measured on the core)house1155%0.54525.2——env-baselines.json, audit 359cf1541a8a…
solver (house, measured on the core)house11100%1.0005.4——env-baselines.json, audit 359cf1541a8a…

Edition 2026-10, 2026-10-16. Scaffold arena/1. Seeds 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10. 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.

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/d44531a693292711f02ba37d4dd92e6f372b1c193e92796df3d6b20d2cfba8c4
python3 tools/arena.py replay d44531a693292711f02ba37d4dd92e6f372b1c193e92796df3d6b20d2cfba8c4