Key answer
Specify protected content and conflict handling together: identify the conflict, propose the smallest change, and let you decide whether to relax the boundary. Discovering a problem does not automatically authorize editing protected areas.
Full explanation
AAAI 2026's ConInstruct tested GPT-4o, Claude and others on conflicting instructions across six task types including email, plans and stories. Models recognized conflicts but often generated answers meeting only some conditions. Asking for conflict detection and clarification increased reporting but also unnecessary questions on nonconflicting tasks. Specify concrete triggers rather than “stop whenever there is a problem.” Sections 5.1–5.2
“Correct every total without changing source records” may require only formulas, or the records may themselves be wrong. For example:
Correct the totals without changing original detail records. If formula changes suffice, proceed. If records must change, identify cells, the conflict and evidence, and propose the smallest fix without executing that part. Continue corrections independent of this decision.
Require three points: conflicting goal/prohibition, why current boundaries make completion impossible, and the smallest permission change needed. Do not accept vague “complete refactoring is necessary” claims without considering retained originals, an accompanying note or a new copy.
Necessary changes differ from improvements. A prettier layout may not be needed to correct data. Only genuinely incompatible hard requirements need a decision.
After approval, inspect differences: changes stay within permission, preserved areas match originals and the error is resolved. Keeping originals and viewing file/code diffs reveals boundary violations better than a completion assurance alone.