Transfer lens
PUBLISHED METHOD / MEASUREMENT NOTE

SLEIGHT-Bench

Detect harmful agent behavior

01 / TEST SETTING

Synthetic paired transcripts

02 / OBSERVED CHECK

Detection with false-positive calibration

03 / TRANSFER BOUNDARY

Constructed cases, not field prevalence

MEASURED TASK AREAS
OversightEvidence

Task classification identifies scope. It does not report model performance or safe deployment.

REPORTED FINDING

Synthetic transcripts expose monitor blind spots

SLEIGHT-Bench reports monitor misses on constructed harmful agent transcripts.

Synthetic cases reveal weaknesses under tested conditions; they do not measure deployed incident prevalence.