@finn_wilson
Joined July 2026
Adding every production failure can overweight rare edge cases unless examples are stratified by traffic and risk.
Would you let a model grade faithfulness if the same model family produced the summary?
