Simulator Trust / Policy gml-heuristic-v1
How AI Plays
Gundam Match Lab uses deterministic rules and heuristic ranking. It does not use a generative AI service, and its scores are rankings rather than probabilities.
Legal actions
Each decision begins by generating actions allowed by the current hand, board, timing, resources, zones, and modelled card requirements. An action that cannot pay its cost or select a legal target is excluded.
Ranking and tie-breaking
Legal actions receive a heuristic score based on board development, removal value, pressure, resource efficiency, and card-specific utility. Equal scores use deterministic card-cost and card-order tie-breakers so the same seed and inputs reproduce the same result.
Mulligans and targeting
Mulligans favour playable early hands. Targets are ranked from currently legal choices using threat, survivability, and effect utility. These choices are practical approximations, not claims about universally optimal play.
Replay decisions
When replay recording is enabled, the simulator may retain the chosen action, up to three legal alternatives, heuristic scores, reason codes, and policy version. Effect requirements, costs, targets, and resolution steps remain in the separate Decision Inspector.
Unsupported behaviour
Cards marked Stats only can still be drawn and used through shared combat rules, but their unique effects are not executed. Partial cards execute only the behaviour listed as modelled. Coverage and statistical confidence are reported separately.
Known limits
The policy does not model bluffing, opponent psychology, open-ended table talk, every ordering preference, or perfect tournament time management. Tournament comparisons are observational unless their deck definitions and samples match the simulator setup.