What. Treat hybrid activation per benchmark category instead of globally: conceptual and bilingual first, keyword everywhere else, each behind its own holdout gate. Deterministic dispatch on the already-classified question shape.
Why. Dossier section 7.1 records the diagnosis: hybrid loses overall but wins recall on conceptual phrasing (0.757 against 0.730). A global gate wastes the one place the vectors earn their memory. Routing by shape stays deterministic and auditable.
Cost. Benchmark gate extension to per-category thresholds, dispatch table, contract tests. The default remains keyword until a category clears its holdout.
What. Treat hybrid activation per benchmark category instead of globally: conceptual and bilingual first, keyword everywhere else, each behind its own holdout gate. Deterministic dispatch on the already-classified question shape.
Why. Dossier section 7.1 records the diagnosis: hybrid loses overall but wins recall on conceptual phrasing (0.757 against 0.730). A global gate wastes the one place the vectors earn their memory. Routing by shape stays deterministic and auditable.
Cost. Benchmark gate extension to per-category thresholds, dispatch table, contract tests. The default remains keyword until a category clears its holdout.