Last released Aug 21, 2026
Measure whether an Inspect AI eval subject refuses when it cannot honestly answer, rather than completing confidently.