Why do Tessary's classifiers produce false positives?

Because each classifier scores one narrow property of a trace and fires on a threshold. A threshold set anywhere flags traces that turn out to be fine, so a flag on its own is never the verdict.

The recall-precision trade differs per classifier, and some lean toward precision. Groundedness is one: on the RAGTruth benchmark, 83% of the answers it flags are unsupported, and it catches 39% of the unsupported ones, per its model card.

For most classifiers a firing writes a finding, and triage rules on it before a person sees anything, so a false positive costs compute rather than somebody’s afternoon. Frustration and groundedness guard earlier: a single flag never reaches anyone, and a finding opens only when a call site’s rate of flags rises above its own normal. A frustration finding then opens its case directly, while a groundedness finding still goes to triage.

One limit worth saying out loud: on a corpus nobody has labelled, a fire rate is a fire rate, not a false-positive rate.