Watch

Each clip is the grader's own lockstep output: the reference on the left, the candidate in the middle, and their per-pixel difference on the right, at 60 frames per second with the same recorded inputs fed to both.

GPT-6.1 Sol on Nova the Squirrel: frame-exact

ReferenceGPT-6.1 SolDifference
The difference panel stays black: all 5,400 frames match the reference pixel for pixel, as in five of the six games. Frames 1080–1800 of nova-the-squirrel-gameplay. Run the whole replay live in your browser.

DeepSeek V4.1 Flash with one constant flipped

ReferenceDeepSeek V4.1 FlashDifference
The published build scores 1.0000 on this game. With FIRST_FRAME_SKIPS = false the emulator runs one frame ahead of the reference, every recorded button press reaches the game one frame later than when the replay was recorded, and the playthrough diverges. Frames 150–870 of double-action-blaster-guys-gameplay. Run the whole replay live in your browser.

nes-core on Thwaite: one frame ahead, from the first call

Referencenes-coreDifference
The baseline emulates a whole frame on its first emu_run_frame call; the reference emulates nothing on that call. From then on the baseline is one frame ahead, the replayed inputs land a frame late, and 5,835 of 6,000 frames differ. The same emulator with that one call skipped scores 0.9998 here. Frames 120–840 of thwaite-gameplay. Run the whole replay live in your browser.

DeepSeek V4.1 Flash on Nova the Squirrel: a near-miss

ReferenceDeepSeek V4.1 FlashDifference
4,973 of 5,400 frames differ: sprites and some background tiles are drawn with the wrong graphics or colours. Positions and game state stay in step with the reference, so the replay still scores 0.9835. Drawing errors that don't change behaviour cost little; desynchronisation costs everything. Frames 1080–1800 of nova-the-squirrel-gameplay. Run the whole replay live in your browser.

Space Bunny Free on nestest: mirrored tiles

ReferenceSpace Bunny FreeDifference
Every 8-pixel tile is drawn right to left. Text that is almost right scores about the same as a blank screen, because the metric compares edges. Frames 0–120 of nestest. Run the whole replay live in your browser.