Admit frozen evaluation inputs by predecision validity, not a passing baseline answer

A frozen-context model benchmark should validate the input prefix independently of the baseline's answer. Requiring the entire baseline conversation to pass before admitting an input can hide precisely the decisions where the baseline is wrong. Keep input validity, independent expected behavior, and observed model output as separate evidence.

This applies to staged tool-calling flows with auditable native prompts and schemas. Bind the reviewed messages/tool menu to a request hash, check preceding decisions and the currently spoken question, and retain every original failure score. Use a source-selected panel for broad comparison and report any post-discovery failure panel separately. A hash preserves an audited input; it does not prove that input or its expected answer is correct. Historical transcripts without historical prompts or audio support qualified regression variants rather than exact historical replication.

Applies when: Comparing models on identical staged conversation inputs while the baseline itself has failures. Keywords: frozen context benchmark baseline selection bias tool calling regression prefix gold audit