Confidence Thresholds For Model Escalation Routing

A model's confidence score becomes useful for routing only after it has been checked against actual mistakes. OpenRouter's escalation guide recommends grading representative answers and grouping them by their reported confidence. If lower-scored groups contain more errors, an application can use that ordering to select requests for another model. The cutoff must come from those observations: a reported 0.85 does not establish an 85% probability of correctness. Changing the model or traffic requires checking whether the relationship still holds.