Arena comparison 001. 44 matched games per checkpoint, 308 common-opponent games, plus 44 separate mixed-model tournament games. Error bars are 95% Wilson intervals over games, NOT training seeds. One training seed per algorithm. In the common-opponent benchmark, each focal agent faces 10 frozen untrained Lana opponents. The direct tournament instead contains all six trained checkpoints and five untrained agents. 100/1000 labels are training-game budgets, not full dataset passes. Dreamer 4* is an unofficial vector adaptation trained offline on MuZero self-play with separate fresh fits at each budget. Unequal optimizer budgets and direct Dreamer policy vs MCTS planning prevent an equal-compute algorithm-superiority conclusion. Slot 11 has a larger collision radius and faster firing; focal seats are balanced. A separate 132-game CPU Lana audit is preserved and not pooled with these GPU results. Cutoffs count zero. See protocol.json and raw games.csv. No continual-learning/retention claim.