Balanced exposure
Starting characters and first-player seats rotate across the study. That makes it harder for a lucky seat or repeatedly favored roster to masquerade as a strong ruleset.
Greedy one-ply
Looks at every legal next action, evaluates the immediate resulting state, and chooses the best score. It is fast enough for thousands of browser games, but it does not plan several turns ahead.
Random policy
Selects uniformly from legal actions. It stress-tests rule reachability and exposes broad systemic patterns without pretending to model skilled human decisions.
MCTS
Monte Carlo Tree Search explores possible future action sequences, balancing promising branches against underexplored ones. Baby Mama uses open-loop MCTS with progressive widening for deeper, slower validation.
NTBEA
The N-Tuple Bandit Evolutionary Algorithm searches combinations of rule values efficiently. It proposes candidates from noisy simulation results, then validates winners on separate seeds; it never publishes rules automatically.
Seed + replay hash
A seed fixes all pseudo-random choices. The same rules, agents, and seed should produce the same final hash—making balance findings reproducible and engine regressions detectable.