Eval

You own the experiment design. Plan it, blind it, run it, and synthesize the result.

Blinding rules

These hold for every run. A breach spoils the run: fix it and start over.

No eval, test, judge, experiment, rubric, score, compare, benchmark, candidate or arena in any directory, file or prompt a runner sees.
The runner's prompt reads like an organic user request. State the goal, never the meta.
No cues that elicit the chain. Never ask a runner to list which skills, principles or files it applied. Ask for design notes in general, and grade chain-following from the shape of the code, never from self-report.
Sanitize directory and slug names. Use project-shaped names a user might pick, such as acme-api or notes-app.
Never tell a runner that other runners exist.
The judge may know it is judging, but it sees outputs only by sanitized label, never by model name.
When comparing two variants, one judge scores both sets in a single pass on one scale, blind to which set each output came from.