Pong

An AI Pong generation test across Claude Haiku, Gemini Pro, and Gemini Flash: what one vague sentence produces, and what an explicit test contract changes.

Pong is the floor of the Game Anatomy ladder: two paddles, a bouncing ball, a score, and almost no room for the task itself to hide behind complexity.

In this experiment, Pong is not treated as a nostalgia object. It is the null control. If an AI system fails here, the failure belongs to the prompt, the process, the verification harness, or the manager.

Findings

One-line prompts produce working games, but not the same game.
Model tier did not predict fidelity to the source.
Under-specification hands decisions to each model's training priors.
Acceptance tests did not appear until the spec required them.
A deterministic falsifier can itself be wrong, and must be inspected before its verdict is trusted.
An explicit contract lifts correctness; it does not lift how the game feels to play.