Reading a First-Run Regression Report

You run all 40 golden sets, unchanged prompts, against the new model. Results: 19 pass with no meaningful change, 9 show format drift (correct content, wrong shape), 6 show tone/style drift, 4 show a genuine capability drop on a sub-task, and 2 show a refusal-behavior change.

1. Explain why lumping all 21 regressions into one "fix the prompt" bucket is a mistake.
2. For each of the four regression types, describe the fix you'd reach for and how confident you are it will fully close the gap.
3. Which of the four types most needs a non-engineering sign-off before you touch it, and why?