Definition
Plain language
A score for how well a detector can tell real signals apart from noise, separate from how trigger-happy it is.
As stated in the literature
The sensitivity index from signal detection theory measuring the separation between signal and noise distributions, distinct from the response criterion (bias); invoked in cross-section defect-detection analysis to argue that alignment training shifts the criterion rather than improving sensitivity.
Why it matters: It matters because it separates a detector's true skill from its mere willingness to raise an alarm, so you don't mistake caution for competence.
For example, two airport scanners might both miss the same fraction of threats, but the one with higher d-prime is genuinely better at telling a weapon from a harmless object rather than just guessing more cautiously.
Heard on the show
“Which we should explain without ever using the phrase d-prime out loud.”Episode 087 — When No Agent Reads the Whole Document: A Universal Cliff in Multi-Agent Review