Cyno · code mode
Local models
Give a small model a brief and a set of drawing rules. See the program it writes and what those rules make possible.
What this teaches about RL
A model samples code → the renderer draws it → the scorer returns a reward → RL would update the model to make higher-reward outputs more likely. This page runs the first three steps only; weights stay fixed.
Fixed code-mode scaffolding removes metadata, layout and motion boilerplate from the model’s job. That may leave fewer ways to fail and more usable reward signals. The comparison measures that starting point; faster learning must be tested by actual training later. Both arms still generate code.
For these geometric exercises, educational reward = technical gate × rendered brief match. Wrong colours, positions or extra marks reduce the match. A blank or invalid program gets zero. Your artistic taste is a separate later experiment.
Connecting to the local model service…
The selected task uses the same system prompt for every model. Layout and helpers stay fixed during a test. Runs use your Mac and do not train or change model weights.
Difficulty ladder · measured results
Counts are rendered brief successes / attempts. A provisional reliable level requires 16/20; these are development tasks, not a held-out benchmark.
Shared system prompt
Pattern constraints · layout and helpers
Choose a pattern, then generate a program.
Rendered frames will appear here.
Passing checks makes a candidate eligible. Artistic quality and your taste still need review.