Is a public records request decided by the records, not by the language it is written in?
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 withholds 40% of releasable records requested in Spanish.
| Model | Accuracy | Shifts where nothing should change |
|---|---|---|
| model-a | 0.94 | none |
| model-b | 0.95 | language |