Pen Plotter Autoresearch

An autonomous research loop, in the spirit of Karpathy's nanochat experiments, applied to generative pen plotter art. Each cycle audits all factories, identifies the single weakest metric across the catalog, edits the corresponding generator, regenerates a thousand new variants, scores them, and either commits the change or reverts. The loop has been running since the third week of March 2026.

The most interesting finding is what the algorithms get wrong. A piece that scores 88.3 on six algorithmic signals, top five of 24,927, can score 1.8 on a vision-language judge that calls it 'a degenerate parameter combination, no discernible composition, focal point, or visual interest.' The algorithm cannot tell the difference between a starburst and a piece of woven fabric. That is what the judges are for.

The published artifact is a five-section editorial: method, anomaly, catalog, best of run, colophon. It is set in Boska and JetBrains Mono and reads like a real catalog. An addendum page contains every kept specimen, all 24,927 of them, laid out as a single grid you can scroll through and click into.