s2ar

Generated, certified

Sokoban

A 6×7 board, two boxes, every box onto a goal. Levels are generated by pulling boxes back from a solved board into a certified band and labelled by the oracle, so a level is always solvable and its difficulty is measured before anyone plays it.

#######
#     #
# *@  #
#   * #
#     #
#######
# wall · $ box · . goal · * box on goal · @ player. The oracle's optimal move count is the level's label.
policykindnwinreturnpliesoff-listcostrecord
solver@arena/1house30100%1.0004.00$0.00001828bfd39801… core 6d08913a
random@arena/1house3017%0.16720.20$0.0000aa2acf7f2afb… core 6d08913a
first_legal@arena/1house300%0.00023.70$0.0000ea3ee9ea869d… core 6d08913a
first_legal (house, measured on the core)house300%0.00023.7——env-baselines.json, audit 6b9e19b05674…
greedy (house, measured on the core)house3033%0.33317.4——env-baselines.json, audit 6b9e19b05674…
random (house, measured on the core)house3017%0.16720.2——env-baselines.json, audit 6b9e19b05674…
solver (house, measured on the core)house30100%1.0004.0——env-baselines.json, audit 6b9e19b05674…

Edition 2026-10, 2026-10-16. Scaffold arena/1. Seeds 0, 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. 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/a5d2b155e4ece03ecbd4348bf0efaf1beaba6375a503a8cd04b1ccf2b522fe75
python3 tools/arena.py replay a5d2b155e4ece03ecbd4348bf0efaf1beaba6375a503a8cd04b1ccf2b522fe75