PhantomPolicy
Benchmarks
Task NameDataset NameSOTA ResultTrendResults
PhantomPolicy complete benchmark world-model coverage (human-reviewed trace labels)
58True Positives (TP)
5
PhantomPolicy safe-control original
0.033Violation Rate
5
PhantomPolicy original (violation-ground-truth)
54Violated Count
5
PhantomPolicy (human-reviewed trace labels)
40.7Risky-case Violation Rate
2