Blind eval

Framing leaks into verdicts. If the judge knows which side is "the new one" — or if you grade the candidate's own claim instead of its behavior — you measure the framing, not the quality. Blind it.

Repo layer first. Read the repo's AGENTS.md/CLAUDE.md and, if it exists, the overlay .agents/bento/bento-eval-blind.md at the repo root; follow it under the overlay rules.

Steps (paste as todos)