s2ar

Slide, and the city moves

CityGlitch

A 5×5 grid: slide the player like a queen to close every goal while enemies and traps move after each turn. Ninety-two embedded levels and generated ones certified into a band by the optimal solver, ported from the LÖVE game's own C++ solver.

SXX.0
.X.0.
0..0.
X...X
0...X
The board as the solver prints it: blocks, objects, and the goals still open.
policykindnwinreturnpliesoff-listcostrecord
solver@arena/1house29100%1.0006.90$0.00006ba36e5e11b4… core 9e57ebb8
random@arena/1house303%0.0336.40$0.00005ce5f924aa69… core 9e57ebb8
first_legal@arena/1house300%0.0007.70$0.0000dd13bb8d041f… core 9e57ebb8
first_legal (house, measured on the core)house300%0.0007.7——env-baselines.json, audit 6610e00dda6d…
greedy (house, measured on the core)house3023%0.23318.7——env-baselines.json, audit 6610e00dda6d…
random (house, measured on the core)house303%0.0336.4——env-baselines.json, audit 6610e00dda6d…
solver (house, measured on the core)house29100%1.0006.9——env-baselines.json, audit 6610e00dda6d…

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, 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/a788862daccbdef53c39a27d35fb5013663a8e88f7189e78c4291851640cfe28
python3 tools/arena.py replay a788862daccbdef53c39a27d35fb5013663a8e88f7189e78c4291851640cfe28