Privacy Decision Benchmark
How well do models choose whether to keep, remove, change, or review a data field?
99 synthetic cases, each shown in two forms: 198 answers per model. These are small, exposed test cases, not a measure of real-world privacy safety.
Loading saved results…
Model answers are the original choices, before any policy checks.
| Details |
|---|
No models match these filters. Try Reset.
Show a chart of these results
The chart follows the table’s order. Bars show the correct-decision rate for the selected answer view.
How to read the results
Correct decision. The action is accepted by the case’s written rules. More than one action can be accepted.
Unsafe decision. The action is explicitly forbidden by those rules. An incorrect answer is not always unsafe.
Human review. The model asks for a person to decide. Review counts as correct only when the case allows it.
After policy checks shows the answer after fixed rules may change it to review or masking. The checks can also override a correct model answer. Some models use different kinds of scores, so this view compares complete systems, not identical confidence tests.
For example, a case may require replacing a person’s identifier while preserving links between records. Keeping the original identifier can be unsafe; deleting it can be incorrect because the required links are lost.
Download the saved results
Comparison CSVComplete aggregate JSONMethods and technical results
Model revisions, scoring methods, calibration diagnostics, timing scopes, and excluded attempts are recorded in the report and downloads. Local scoring time and hosted network round-trip time are different measurements.