reviewers have to earn it

been reading axelrod. the evolution of cooperation.

so i made my ai reviewers play it. three models grade my drafts. now they have to earn the right to be trusted.

every round i log which flags were real and which i rejected, with a reason.

cooperate and your reviews get lighter. flag noise and you get read harder.

one model runs 0.91 precision. one runs 0.58 — half its flags just prescribe away my voice.

turns out review isn’t about how many eyes.

it’s which eyes have been right before.

tit for tat, but the players are the models grading you.