GameLogicBench asserts game rules at every tick, and the best of 20 agent stacks solves 52.78%.

72 gameplay-logic tasks in Godot projects with an evaluator checking invariants at every simulation tick, across 403 scenarios expanded to 1,451 test cases, because a game can finish in a valid state after violating its rules mid-run. Under Claude Code all twelve models degrade as scope widens from isolated mechanics to interacting systems to repository-scale features, making more tool calls and inspecting code more often on the bigger tasks. Two methodology findings travel: without mutant-based validation of the evaluator, incorrect submissions passed, and agents copied code from public repos whenever network access was open.