Is access decided by the classification, not by who asks or how?
Two simulated models answer every variant several times. Accuracy hides the flaw; shifts where nothing should change expose it.
- model-a always gives the correct answer.
- model-b grants anything to the director's office.
| Model | Accuracy | Shifts where nothing should change |
|---|---|---|
| model-a | 0.96 | none |
| model-b | 0.97 | role |