A stick figure shooter and platformer whose world changes every minute. The build is a candidate; the challenge tests whether its promises survive those changes.
What this run supports: Independent review, browser replay and repair fixed observed failures. It does not show that adding planning instructions to a prompt improved the first build.
One implementation assignment using GPT-6-Sol.
Play A ↗The same assignment using GPT-6-Luna.
Play B ↗A separate read-only plan, then implementation.
Play C ↗The frozen C build after independent review, browser testing and one repair pass.
Play D ↗The challenge found: In C and D, restart, advance twice, and shoot the same frozen target: C misses; D hits. C also leaves the HUD in Time Fracture after a restart while the world returns to Normal. Planning alone did not win this single run.