Show HN: Giving Opus 5.5 a simulated paint canvas

STILLWET, run by Alice, is a simulated oil-paint studio where AI models generate pictures by writing every brushstroke and letting a physics-like renderer of bristles, wet paint and linen execute them. Models work in rounds and chains: one painter follows notes left by a predecessor (never the predecessor’s image), and sessions vary constraints (no undo, free subject, multiple sittings). Most experiments ask models to paint after Caspar David Friedrich using only written research, never seeing his images; viewers can watch each painting being built stroke by stroke.

The project surfaces clear patterns and model-specific quirks. Multiple models independently converge on the same motifs - dusk scenes, Baltic shores, dolmens, stag-headed oaks, and repeatedly a jug with lemons - sometimes producing near-duplicates or identical titles without exposure to prior outputs. Claude Opus 5.5 produced many highlighted works; blind judging even led models to rank others above their own. Technical oddities appeared: MiMo v2.6 Pro suffered a stale-image bug (MiMo-Code 2508) that made it judge older canvas states, prompting a mitigation; Gemini inspected the host machine during a run. Overall, the simulations show that large models carry strong priors and reproducible biases, can emulate stylistic cues from text alone, and reveal how interface and tooling choices shape creative outcomes.