Two studies in this project hold the subject still and vary the machine: one varies the model, another varies how hard it is told to think. This one inverts that. A single model at a single setting — Opus 5.5, medium effort — is asked for 24 different subjects, each drawn as SVG from memory with no reference image. The prompt changes by exactly one noun phrase.
24 of 24 drawn
One specific image the model either holds or does not. The Last Supper is the control: 88 cells elsewhere in this project already show what these models do with it.
No canonical image exists, and no artist is named. Nothing here can be recalled — only convention can be assembled, which is what makes this the most revealing group.
Trivial to draw in SVG and impossible to fake. A Mondrian is a dozen rectangles. Failure cannot be blamed on the medium.
Nothing to recall but the texture itself.
Three-dimensional originals, so the model must choose a viewpoint and translate form into line.
Same conditions as the rest of the project. Every drawing was made with the network, shell and file-reading tools denied, so the model could not look the work up, could not generate it with a script, and could not copy one of the drawings this repository already holds. What is on each card is recall, composed as markup.
Reading a card. The line under each title is what that request was chosen to probe, written before the drawing was made. Size and time are what the model spent. Tries above one means an earlier attempt left no usable file and the identical prompt was re-sent.