Your team is deciding between always retrieving and bui...

The ‘below-closed-book’ finding suggests that retrieval can sometimes hurt performance if the model is forced to process irrelevant or distracting context. A router can mitigate this by skipping retrieval for queries where the model already has sufficient parametric knowledge.

To justify building a router, you must measure the net-negative zone for different query types. Specifically, you should calculate the change in accuracy (∆p) based on document position across a representative evaluation set. You should prioritize routing for query classes that exhibit high parametric confidence, as these are the cases where the model has the most to lose by being forced to attend to retrieved context.