Freeze the harness before attributing tool-routing errors to the model

When a tool-routing boundary is explicit in both system instructions and tool descriptions, and both a denser prompt overlay and forced tool choice worsen semantic false positives, treat the remaining error as a model-and-data contrast problem before rewriting the harness again. Keep the original harness fixed and compare one contrast-trained candidate with the current model on a freshly frozen, disjoint holdout. Report each direction of the boundary and protected-route regressions separately.

A later harness arm should change only one description and remain diagnostic unless it improves both directions without new false captures. This stop rule does not prove that every explicit prompt is adequate, and a source-derived evaluation contract does not establish parity with deployed wire behavior.