Particle 1 added

Particle-1 produced HTML for all 35 tasks, with every session ending normally and a total recorded API cost of $4.12. Three transcripts offer a closer look at how it tested its work: temporary code made game state and animation behavior measurable, and the returned results sometimes changed the next step.

2048 contains a clear repair sequence. Particle-1 inserted ten checks using preset boards. Nine initially failed: merging a row of four 2s returned stored values of 2 and 2, even though the score gain was correctly eight. It updated the merged tile’s stored value, then reloaded; all ten checks returned PASS. Further checks exercised the win overlay, continuing play, a full board with no moves and New Game. It removed the temporary test code before the final reload. Start here to follow a bug from a failing check through a specific edit to a passing result.

Maze Generator goes beyond watching a solver animation. A separate browser script inspected one generated 25-by-15 maze: all 375 cells were reachable, with 374 connections, and the reported solution was contiguous from entrance to exit. Another check interrupted the solver with New Maze; the explored-cell count and saved solution both reset to zero before generation restarted. These are recorded checks of particular generated mazes, rather than a guarantee about every random layout.

Matrix Rain shows why the test itself also matters. Its first pause test compared entire screenshot files and reported a mismatch. A follow-up script measured the canvas directly: the sampled pixel hash and frame counter stayed identical for one second, then changed after Resume, without another application edit. The same session checked six viewport sizes and a double-density canvas. Those observations support specific resize and control behavior; the live build lets you judge the visual effect yourself.