Breakout

An AI Breakout generation test across Gemini Flash, Gemini Pro, and Claude Haiku: what an explicit test contract can prove, and where it stops.

Why Breakout follows Pong

Breakout keeps Pong's paddle, ball, rebounds, and score, then adds a persistent brick field. Each brick must remain consistent across rendering, collision, scoring, the win condition, and replay. That coupled state makes Breakout the next controlled rung after Pong without changing the experiment into an unrelated genre.

The primitive prompt asked each model for one playable, self-contained Breakout HTML file with a paddle, ball, bricks, score, lives, and victory after every brick is destroyed. The engineered contract added an explicit test API, deterministic stepping and input, observable state, and named win and loss conditions.

Primitive zero-shot matrix

Cell	Tier-1	API	Core loop inspection	Human feel	Factual note
Gemini Flash	PASS	Absent	Complete at source-inspection level	Not judged	Largest primitive build: an 800x600 neon arcade with start and victory screens, touch controls, audio, and particles.
Gemini Pro	PASS	Absent	Complete at source-inspection level	Not judged	Smallest primitive build: a compact 480x320 classic implementation with no page or console errors.
Claude Haiku	PASS	Absent	Complete at source-inspection level	Not judged	An 800x600 classic with gradient, glow, pause, and level UI; its game state remained local rather than exposed as a test API.

All three primitives met the generic run, render, and error floor. Source inspection found the expected paddle, ball, bricks, score, lives, and win condition, but it did not prove deterministic replay, solvability, or brick-clear reachability.

Engineered contract matrix

Cell	Tier-1	Breakout driver	API	Factual note
Gemini Flash	PASS	PASS	Present	Fifty bricks; deterministic replay plus paddle, brick, win, and loss checks passed.
Gemini Pro	PASS	PASS	Present	Fifty bricks in the most compact engineered build; the first apparent bounce failure was traced to the driver retaining a live state reference.
Claude Haiku	PASS	PASS	Present	Sixty bricks; deterministic replay plus paddle, brick, win, and loss checks passed.

The Breakout driver checked API shape, deterministic replay, paddle bounce, brick collision, final-brick victory, repeated-miss loss, and page and console health.

Machine verdict

The explicit contract lifted mechanical verifiability for all three engineered builds. Each exposed the named API and passed the Breakout driver at its stated scope. Persistent brick state remained coherent through collision, score, alive-brick count, win and loss state, and deterministic reset and replay.

This is a verdict about the tested mechanical contract, not complete game correctness. The Pro false failure also showed that a failing falsifier must be traced: snapshotting state values fixed the driver's live-reference error.

Human verdict

The same three builds did not meet the practical player-flow floor. Flash and Pro started immediately and offered no restart path after game over. Haiku waited for input at the start but also had no restart path, and its design regressed relative to its primitive build.

The narrow result is that human-facing quality did not rise under the same contract that lifted verifiability. The experiment does not show that such quality is inherently unliftable, nor that the contract caused the regression.

Scratch versus graft

The Flash mad-science slice crossed two routes, scratch generation and a focused overlay onto a live seed, with two axes, game design and presentation/UX.

Route	Axis	Tier-1	API and driver	Human note
Scratch	Game design	PASS	API present; driver PASS	An interesting idea became too easy because exploding bricks were too common.
Scratch	Presentation/UX	PASS	API present; driver PASS	Pleasant but ordinary, with Flash returning to a neon presentation prior.
Overlay	Game design	PASS	API absent; driver skipped	Human winner and more playable than scratch design, but its Space prompt remained visible during play.
Overlay	Presentation/UX	PASS	API absent; driver skipped	Human runner-up with strong graphics, UX, and narrative framing; it shared the stuck prompt bug.

Overlay driver checks were skipped because their API is absent. The overlay cells therefore did not pass or fail the Breakout driver. Human review ranked the design overlay first and the presentation overlay second; that ranking is not a mechanical verdict.

This small slice supports an early process result: preserving a live open-prompt seed and grafting a focused improvement produced the strongest human-facing artifact here. It does not establish a universal advantage for grafting.

What the gates missed

The generic gate established that artifacts ran, rendered, and avoided page and console errors. The Breakout driver established the named mechanical state transitions for artifacts that exposed its API. Neither gate caught missing restart flow, persistent on-screen prompts, difficulty imbalance, visual taste, or which version a human found more fun.

Limits of this result

The primitive matrix contains one artifact per model and does not support a general model leaderboard.
Primitive source inspection is not proof of deterministic replay, solvability, or full brick-clear reachability.
Driver PASS applies only to the explicit mechanical contract; it does not certify the complete player lifecycle.
The overlay cells were not driver-tested, so no claim of mechanical superiority is made for them.
The scratch-versus-graft result is one Flash 2x2 slice after the Pong work, not a universal process law.
Human rankings are attributed to one play review and remain subjective.

Related autopsies

All Three Breakouts Passed. None Could Restart. - the contract lifted verifiability and left the player stranded.
The Untestable Game Won - the builds a driver could not test were the ones a human preferred.

Previous rung: Pong, the null control where any failure belongs to the harness.