Retrieval Precision

Precision and recall trade against each other, and the balance appropriate for consumer search is wrong for governed retrieval. Low precision fills a context window with marginal material, crowding out governance instructions and increasing cost per answer while degrading quality — the model has more to reason over and more opportunity to draw on the wrong passage. High precision with modest recall is usually the better setting in regulated work, because an answer that cites three correct sources beats one that surveys twenty of mixed relevance.

Taxonomy scoping in Centralpoint raises precision structurally rather than by tuning similarity thresholds: a prompt bound to a branch of the hierarchy retrieves within it, so marginal material from adjacent domains is never a candidate. SkillTokenBudget then bounds what the execution may spend, which makes low precision visible as cost rather than as a quiet quality drift.