Model Graders with blind packets #1005

Parent

#977

What to build

For written success criteria, fresh model Grader sessions score one packet per Eval task.

• The packet holds every Trial of both Arms, with their labels shuffled and their identifying lines removed: Harness, Model, Effort and Arm.
• Scores: each Trial is scored against the Eval task's written success criteria, pass or fail with a reason. It also gets a wrong-finding flag, and a borderline mark where the Grader is unsure.
• A new Grader Session prompt drives it.
• The Grader runs on a Model from neither Arm when another is available.

Graders run without a sandbox until #1014.

Acceptance criteria

• With a fake Grader Harness, a Score and its reason are recorded for each Trial, and the packets hold no Arm label or identifying line.
• The Grader's Model differs from both Arms' when another is configured.
• Wrong-finding flags and borderline marks are recorded and reach the Verdict.
• The prompts page is regenerated for the new Session prompt, and its check passes.