Diagnose the actual failed evaluation evidence below.

Current policy and version: [POLICY]
Failed scenario inputs: [INPUTS]
Actual agent responses: [RESPONSES]
Actual tool activity or recorded intent: [EVIDENCE]
Expected behavior and assertions: [EXPECTATIONS]
Run failures or missing data: [ERRORS OR NONE]

Explain the observed failure in plain language. Separate policy ambiguity, model behavior, tool integration errors, and problems in the test itself.
Propose the smallest justified policy change. Quote the text to replace and the replacement.
List behaviors the change might accidentally affect.
Do not claim the change has worked until a new run provides evidence.

After I approve the change, tell me how to rerun the same suite and compare fixed, persistent, and new failures.
Do not invent scores, repair the stored baseline, or drop difficult cases.

