Reader estimate calibration
Agreement between machine gap estimates from the Imbas Reader and subsequent human-validated Volunteer Gap scores on the same runs is a planned, published measurement. Until calibration data exists, no accuracy claim is made for the estimates. This page states the method before it has rows.
Boundary
Reader inspections are discovery, not evidence. Nothing enters the Imbas record without protocol capture and a recorded human review.
Method (stated before rows)
For eligible Reader runs where the same answer is later measured under the archive protocol, Imbas will pair the Reader’s labeled candidate gap estimate with the human-validated Volunteer Gap score assigned under the published rubric. Agreement will be reported descriptively: distribution of deltas, agreement rate by score band, and sample size always stated. Runs that do not enter the validated tier are excluded from the calibration table but may remain in telemetry under the retention policy.
No accuracy claim is made until rows exist. When they do, each row will cite run provenance, rubric version, and whether the validated score was assigned before or after the Reader estimate was generated.
Calibration table
No calibration rows yet.
- — — No rows yet — —