drawing domain

Models are told to draw — as raw SVG markup. No canvas API, no image model, no tools. Just a text model forced to produce code that renders into a picture.

why this is a real benchmark and not a party trick

Drawing an SVG from a one-shot prompt stresses several things at once:

• spatial reasoning — a snowman is three circles stacked, a bicycle frame connects the wheel hubs. Getting coordinates coherent is genuinely hard for text models.
• structured output under constraints — the prompt is a contract: exact element types, minimum counts, required attributes, forbidden elements.
• instruction compliance — "output ONLY the SVG, no code fences" filters the chronic over-explainers.

how it's scored

Same rules as everything else on slopmark: no LLM judges. The html_contract verifier parses the output structurally:

• contains_tag — the right shapes exist in the right quantities (<circle> x12 for the smiley grid)
• attribute_exists — shapes are actually filled/stroked, not invisible
• required_substring — viewBox present, specific colors used where the task demands (fill="red" on a traffic light)
• forbidden_pattern — no <script>, and for pixel-art tasks no cheating with <path>

Score = fraction of rules passed. Pass = all rules pass. The gallery on the challenge page renders every drawing exactly as the model coded it (sanitized), so you can also judge with your eyes — the verifier checks anatomy, not beauty.

the harness

Zero context mode: no system prompt, temperature 0. The only difference from other domains is max_tokens is raised to 4000, because a 12-rect pixel heart doesn't fit in 600 tokens and reasoning models spend tokens thinking before they draw.

drawing contest v1

The seed session: 5 Fireworks serverless models × 10 drawing tasks, from "a simple house" (easy) to "sunset over mountains with a linearGradient" (hard). Results live at /challenge/drawing-contest-v1.