Learning lab

The policy learns which planned bridge to build next and which junction to sabotage. Claims races also offer isolated claims, blocking bridges and branches. The flight race adds three shared flights per team and learns when to spend them. Legal routes, movement, interception and retreat still use the existing rules. Each iteration updates the policy from eight new matches. Three quarters of training opponents are frozen versions of earlier policies; the remainder use the scripted bot.

Every checkpoint is evaluated against the scripted bot with both team colours. Score is wins plus half the draws. Actual training replays record the last episode before each checkpoint update, including losses and timeouts. Their seeds change naturally with training. The fixed comparison test keeps the same seed and scripted opponent for every checkpoint; its matches may look identical. Older checkpoints only have fixed tests. A higher training score alone does not prove a more interesting game.