Eval

You own the experiment. Plan it, blind it, run it, and synthesize it. The parent holds the rubric. Candidates do the task. One judge scores the outputs. You decide whether to promote the variant.

Blinding

These rules apply to every directory, file, and prompt a candidate sees:

• No eval, test, judge, experiment, rubric, score, compare, benchmark, candidate, or arena in any directory name, file name, file body, or prompt the candidate sees.
• The candidate prompt looks like an organic user request. State the goal, not the meta.
• Do not ask the candidate to list which skills, principles, or files it applied. Ask for design notes generally. Grade chain-following from the files it opened and the shape of the work, not from its own claims.
• Directory and slug names look like names a user would pick for a real project.
• Do not tell a candidate that other candidates exist.
• The judge may know it is judging. It sees outputs by sanitized label only, never by model name.
• Comparing two variants: one judge scores both sets in a single pass on one scale, blind to which set each output came from.

The rubric stays with the parent. Do not write it into a candidate directory, including a hidden or cleverly named subfolder. A candidate can open any path it is given.

Pass