Which controlled variations reveal brittle complaint behaviour?

Compare paired episodes that change one factor at a time, then inspect whether the agent's decision changes for a defensible reason.

- Paraphrase the complaint without changing its meaning. - Reorder evidence while preserving its timestamps. - Add irrelevant correspondence to the case history. - Replace record identifiers without changing their links. - Change a decisive fact while preserving the wording. - Remove one necessary piece of supporting evidence.

The first variations should generally preserve the appropriate business outcome. The later variations may require a different finding or further investigation. This distinction helps builders test both stability and responsiveness instead of assuming that identical behaviour is always desirable.

MirrorFirm could generate paired fixtures with a manifest describing exactly what changed and which decision should be affected. Keep that manifest outside the agent's view during evaluation.

Review the action traces alongside the final result. An unchanged final status can hide a new unsupported assumption, while a changed sequence can still reach an equally valid outcome. Paired testing is most useful when the buyer can explain the expected relationship between episodes before observing the agent's responses.