Building an unbeatable tic-tac-toe, the spectastic way

Before a line of code, five non-negotiables. The load-bearing one — "rules are pure and framework-free" — is why the board, legal-move check, and win detection ended up as plain functions with no DOM and no AI, the bedrock every engine and every test could stand on. Another — "correctness is demonstrated, not asserted" — is what later forced a self-play harness instead of a hopeful comment.

What to build, for whom, and how we'd know it worked — before deciding how. Ten requirements pinned it down: only legal moves (FR-001), immediate win/draw detection (FR-002), six engines behind one interface (FR-003), two modes (FR-004). The interesting part is the success criteria: SC-001 — "a perfect engine never loses" — a target written as a test, not a wish.

The how — and where the one real decision lived. Three-player tic-tac-toe only makes MaxN and Paranoid interesting, but a bigger board would blow the move-time budget. Decision D-001: keep the 3×3 board for three players and accept that strong three-way play tends to draw. Every design fact was grounded against reality first — including the measured pruning win below.

The plan, sliced into eighteen ordered, checkable tasks. The pure rules core lands first (everything depends on it); the engines stack up from Random to MCTS; the final phase is the verification harness. Tests-first where it counted — T-110 Minimax carried its own acceptance: never loses.

Drain the tasks, tick the boxes, and let the harness have the last word. The first run failed — a turn-alternation bug let a random player beat "perfect" Minimax once. The test caught what a demo never would; one line fixed it; the rerun went clean. That's the point of writing the target as a test.

Then, after acceptance, a real one slipped through: the board's cells went rectangular as tiles landed. No silent hotfix — it went through triage. Reproduced with measurements (rows at 135 / 127 / 70 px), root-caused to content-sized grid rows, classified at the implementation layer, fixed by making each cell square, and locked in with a geometry assertion so it can't return. The whole trail lives in the triage-log.

The lifecycle isn't only greenfield. Three-player games kept ending in draws, so a change was proposed — larger boards, shorter win lines — with typed deltas and an adversarial risk pass (beyond 3×3 the search turns heuristic; a first-cut 2.1 s move was tuned back under budget). Approved, applied into the spec at v1.1.0, and implemented. A 5×5 board won on three in a row now drew 0 of 8 games.