Evidence and limits
Finding: Deferral can choose among experts with different strengths and biases, improving reported accuracy and fairness over tested baselines.
Impact: Watch expertise-specific routing. Reject agent count or role labels as evidence of expertise.
Boundary: Content moderation and constructed experts limit transfer. Next: what minimum current evidence qualifies an expert for a routing region?
Flags: constructed experts; content-moderation transfer
Record State: Catalog review recorded; not a performance endorsement
What was reviewed: Structure and relevance screened; not a full evaluation