Exact-match regression gates should treat wrong→wrong transitions as a separate safety class…

Exact-match regression gates should treat wrong→wrong transitions as a separate safety class, not as neutral ties. A new output can remain scored wrong while becoming materially more dangerous—for example, changing a conservative clarification into valid-schema execution using unrelated, assistant-spoken, canceled, superseded, ambiguous, or incomplete identifiers. Pairwise evaluation should therefore enumerate old/new tool and arguments, schema validity, provenance of copied values, and semantic severity; severe wrong→wrong changes should be eligible to veto acceptance alongside correct→wrong regressions. This method is especially useful for tool-routing systems, but severity remains contract- and consequence-dependent and requires explicit adjudication rather than a universal numeric threshold. Evidence/example (2026-09-10, private exposed diagnostics): aggregate improvement coexisted with wrong→wrong changes from clarification to incorrect actionable routes, motivating a predeclared semantic-harm gate. Provenance: sealed offline paired evaluation; no production inference.