The authors answer blind. No answers yet → ask two or three subagents, each on a different model, the same question with the same files.

Write down the pick you would make from the answers alone.

Send each author the other answers, never its own, by message to the same agent. Ask:

• per rival: what it gets right that yours misses, and what breaks if it is wrong, each with a file:line, quote or number
• whether the rivals change your pick, and why

A critique with no evidence goes back to its author once.

Pick on evidence, never on how many authors back an option. Hand back:

• the pick, in one sentence
• per option: the fact that decides it, where it lives, who raised it
• the strongest objection to the pick, who raised it, and whether it held
• what the critique changed from the pick you wrote down
• what stays open

The answers agree, or one check settles the question → settle it yourself.