FAILURE IS ALL YOU NEED · CROSS-GAME EXPERIMENT
Asterix: collect gold, avoid enemies. Original MinAtar rules and native sticky actions remain intact. We start from the published Astra-authored network and train on legal corrective trajectories.
Green is the player, gold is treasure, red is danger. Each game uses the same starting seed. Improved inference appears only after the final audit passes.
Development and selection gate seeds are reused. The final audit is separate and runs after selection ends. This is a 1,000-step greedy-policy pilot, not a reproduction of the upstream 10,000-step sampled-action score. Base weights: Rikiya Takehi and collaborators, CC BY 4.0. Game: Kenny Young and Tian Tian, MinAtar 1.0.15, GPL-3.0.