GPT-6.1 Terra added

GPT-6.1 Terra produced HTML for all 35 tasks. Its Asteroids transcript shows how it checked behavior beyond the initial browser view: after shortening the playfield, it ran two JavaScript test scripts with mocked browser objects. The assertions covered steering, thrust, firing, screen wrapping, asteroid splitting, collisions, restart and the next wave. These were targeted logic checks using constructed game states; rendering still needs the live build.

The paired Wolfenstein-style shooters show two different debugging paths. Terra ran a separate browser script with timed key holds and logged player coordinates. Forward movement changed x from 2.5 to about 3.71; holding backward against a wall left x unchanged on the next check. It also inspected shooting and restart state. Sol worked through browser feedback: a blocked mouse-capture request led to removing pointer lock, and it added immediate Space-key firing. Later observations showed ammunition decreasing and a hostile being eliminated.

2048 gives another useful contrast. Terra played a short keyboard sequence and changed the message logic so reaching 2048 takes precedence when the board also has no moves. Sol used a JavaScript test environment with seeded boards to check pairwise merges, scoring, undo, win/loss detection and reset. Scripted verification appears in both models’ records, with different checks on each task.

Sol’s examples were saved on October 5, before later revisions to the task and agent instructions. Read the paired builds and transcripts with those differences in mind: they show what each session actually checked, while the live games let you explore behavior beyond those recorded checks.