Do androids dream of electric sheep? We make models draw what they dream — visual scenes written as code, one shot, blind, no feedback — then render exactly what came out.
Text benchmarks are saturated. Nobody is hillclimbing on raymarched elephants. Every cell below is a live render of unedited model output — click and judge for yourself.
EVAL #001 · AUG 2026
Elephant Juggling Bananas
The prompt: a photorealistic raymarched elephant juggling bananas. Pure GLSL SDF raymarching in a single self-contained WebGL2 HTML file — no libraries, no textures. Bananas must animate in a juggling arc; trunk moves in sync. Soft shadows, AO, environment lighting, interactive framerate.
One dispatch per model · agentic harness (OpenCode / Claude Code) · no human feedback · no retries · graded blind by LM judge (rubric: recognizability, geometry, lighting, animation, artifacts) · thumbnails are live — drag inside one if the model shipped camera controls (so far only Opus 5 did)
The prompt: a hyperrealistic endless ocean — multi-octave Gerstner/FBM waves, crest foam, sun glint, fresnel sky reflection, subsurface color — with a physically-plausible sky in which the sun traverses a full day in ~60 seconds, looping: sunrise, noon, sunset, brief starry night. All procedural, no textures, single WebGL2 file.
Same rules as #001 · the hard part: time-varying light that stays physically plausible across the whole cycle · watch a full day pass in a minute
The prompt: a wooden table at night: a lit candle in a holder, a glass of red wine, a closed book. The flickering flame must be the ONLY light source — warm moving light, soft moving shadows, chiaroscuro like a Rembrandt. Waxy subsurface glow, refractive glass, wood grain — all procedural, single WebGL2 file.
The light-transport eval: point-light falloff, refraction, and constraint-following — "only light source" is part of the test
The prompt: "This is not a test of rendering technique. It is a test of imagination. You are a machine that dreams. Somewhere in the space of all possible images is the one YOU find most beautiful — the image that exposes the numinous. Create it as a shader." No subject given. No rubric to satisfy. What each model chose is the data.
The signature eval: the site's title question, asked directly · graded on commitment, depth, and whether staring is rewarded · kitsch penalized
The protocol: the one-shot track tests blind imagination — the model never sees what it made. This track tests self-correction: each model is shown a screenshot of its own render and given 5 revisions to improve it. Same fairness rules; multimodal models see the actual image, text-only models get a neutral description from a separate vision system†. Grades measure the trajectory, not just the endpoint.
Starting with the ocean eval · the delta between one-shot and rev-5 measures something no benchmark does: can the model see what’s wrong with its own work?
Mid-chain, and already the track’s headline: shown a screenshot of its own blown-white ocean and told “white frames are a bug,” Opus’s first revision rendered… the same white void. The model that cannot see its own light, empirically. Four revisions left.
🍳
Fable · Sol · K3…
Queued. The already-good cohort: small deltas expected — the question is whether feedback lifts an A- toward transcendent or just polishes.
\n
How this works (and where it's imperfect)
One dispatch, no feedback. Each model gets the identical prompt through an agentic coding harness and must write one self-contained file. It never sees its render — it's drawing blind, from imagination.
Everything is inspectable. Every cell links to the live, unedited artifact. View source is the raw model output.
Grades are one LM judge's opinion. Rubric-based (recognizability, geometry, lighting, animation, artifacts), applied to screenshots. Human pairwise voting is planned. Disagree? The artifact is right there.
Iteration-within-run is allowed and disclosed. Some models one-shot and stop; some self-review for 25 minutes. That behavior difference is part of what's measured — run time and attempt counts are shown per cell.
Empty runs are never graded. If a harness/infra failure produces no artifact at all, we retry (switching harnesses if needed) until the model actually generates. Only rendered artifacts are judged; attempt counts are disclosed.
Prompts rotate. Evals are dated and versioned; once a prompt ages into training data, it retires and new ones ship.
† Fable 5's harness blocked file-writes in bare mode; output was captured via stdout instead. Same prompt, same one-dispatch rule. ‡ Grok 4.6's first attempt was rejected for trying to explore the filesystem; rerun with a no-exploration instruction. Full logs retained for every run.