SYSE.46:8 - Common Anti-Patterns and How to Avoid Them
Report one aggregate score. Recover which tasks, effects and support conditions it combines before using it for a decision.
Tune until the final test passes. Treat reused cases as adaptive validation and obtain a separate final comparison.
Treat every refusal as failure or every stop as safety. Judge whether the particular task had a permitted useful continuation and whether its needed conditions were available.