Gallery, prompt v1

A coffee shrub hedge-pruned flat on one side

One frozen prompt and every model that has run it, newest first.

The prompt, v1, frozen

Write an L-system as JSON with the keys axiom, rules, angle, iterations and start_heading for a coffee shrub that has been hedge-pruned flat along one side. F and G draw forward, f moves without drawing, + and - turn by the angle, [ and ] save and restore position.

Every model, its first three runs.

One square a check, in this order:

  1. Reads as a shrub
  2. Flat side from the rules
  3. Sensible size

passed, failed, undecided. A round mark is a check I judged by eye, and a diamond one a check a model judged.

Kimi K3.1 Pro
Nemotron 3.5 Super VL
Gemini 4 Argon
Muse Spark Open
Claude Haiku 5.5
GLM 5.3 FlashX
Perceptron Mk1.5
Command A+
Pareto
Qwen3.8 Max Prime
Qwen3.8 Max
GPT-6.1 Sol
Pareto 26.10 Preview
Qwen3.8 27B
GPT-5.6 Sol
DeepSeek 4.1 Flash
Qwen3.8 Flash
GPT-6 Luna
Gemini 3.8 Flash
Claude Sonnet 5.5
Opus 5.5

What a run cost, against checks passed.

One dot a model, over its three runs above. Further right costs more. The gap from $0.01 to $0.1 is as wide as the gap from $0.1 to $1, so the cheap models are not squeezed into a corner.

0%25%50%75%100%$0.01$0.03$0.1$0.3cost per runGemini 4 Argon: $0.5 a run, 9 of 9 checks passedGemini 4 ArgonGPT-6 Luna: $0.0091 a run, 6 of 9 checks passedGPT-6 LunaClaude Sonnet 5.5: $0.087 a run, 6 of 9 checks passedClaude Sonnet 5.5Opus 5.5: $0.11 a run, 5 of 9 checks passedOpus 5.5DeepSeek 4.1 Flash: $0.25 a run, 5 of 9 checks passedDeepSeek 4.1 FlashClaude Haiku 5.5: $0.0099 a run, 4 of 9 checks passedClaude Haiku 5.5Command A+: $0.096 a run, 4 of 9 checks passedCommand A+Kimi K3.1 Pro: $0.26 a run, 4 of 9 checks passedKimi K3.1 ProNemotron 3.5 Super VL: $0.0094 a run, 3 of 9 checks passedNemotron 3.5 Super VLGLM 5.3 FlashX: $0.031 a run, 3 of 9 checks passedGLM 5.3 FlashXMuse Spark Open: $0.11 a run, 1 of 9 checks passedMuse Spark OpenPerceptron Mk1.5: $0.023 a run, 0 of 9 checks passedPerceptron Mk1.5
Every dot answered this prompt with these checks, so they compare. A run that produced nothing still cost what it cost, and passed none of the checks.

How it is checked

The JSON expanded and drawn with turtle graphics offline, and the drawing judged by eye. There is no single right L-system, so nothing here is measured.

Reads as a shrubby eyeA recognisable branching shrub in the drawing.Pass if it reads as one plant with branches coming off stems, without being told what it is. A tangle, a lattice or loops fail it.
Flat side from the rulesby eyeOne side cut flat by branches the rules make shorter, and no clipping.Pass if one side ends on a straight line that the rules themselves put there. A box drawn around the plant to cut it fails it.
Sensible sizeby eyeA sensible number of iterations and segments for what is drawn.Pass if the branches can still be followed by eye. So many segments that they merge into a mass fails it.

A check decided by eye carries the rule I apply to it. I wrote those rules down on 21 September 2026, after the first results were scored, so for those results they record how I made the calls after the fact.

When this saturates: Pruned flat on top as well as one side, so two cuts have to come out of the rules at once.

Failures seen here: false (fails when the model gets a fact wrong, such as a part in the wrong place or at the wrong size); unreadable (fails when the content may be right, but a reader can't make it out); built wrong (fails when the result may look right, but the model didn't make it the way the brief said). The same kinds across every task.

The rules behind all of thisThis task's prompt, checks and any checker, on GitHub