You test Minimax correctness by validating it on small, fully solvable games and by adding invariants that catch perspective, state, and pruning bugs early. The fastest way to gain confidence is to start with a game where you can compute ground truth: tic-tac-toe, small connect-four boards, or any toy game with a manageable state space. For those, write a plain Minimax without pruning and compare its outputs to your optimized version (alpha-beta, transposition tables, move ordering) at the same depth. They must match in chosen move and returned score for deterministic trees.
Implementation bugs usually cluster in a few areas. First is state mutation: make/unmake move must restore every part of the state, not just the board. Add assertions that a state hash before make/unmake matches after unmake. Second is evaluation perspective: decide whether your evaluator returns “good for side to move” or “good for a fixed side,” and enforce it consistently. Negamax helps because it makes perspective flips explicit, but you can still get it wrong if you mix conventions. Third is alpha-beta logic: incorrect updates to alpha/beta, or wrong cutoff conditions, will cause either missed pruning (slow) or incorrect pruning (wrong results). Fourth is terminal detection: wrong win/loss/draw detection silently corrupts the tree.
A practical testing plan: (1) unit-test move generation for legality and completeness; (2) unit-test terminal state detection with known boards; (3) property-test make/unmake with random move sequences; (4) compare plain Minimax vs alpha-beta on random positions at low depths; (5) test transposition table hits by constructing positions known to transpose; and (6) add deterministic logging at the root: list each legal root move and its returned score. That root move score table is your best debugging tool because it tells you whether the decision changed due to evaluation, search depth, or caching. If your evaluation includes external data (retrieval), lock it down in tests: use fixed fixtures rather than live queries. If your system uses Milvus or Zilliz Cloud during evaluation, record retrieval results for test cases and replay them deterministically so you can reproduce bugs and ensure Minimax changes reflect code changes, not shifting retrieval outputs.