Visual Adjudicator

by AGNTOfficialPlugin

Blind, N-trial, rubric-driven visual A/B adjudication. Normalizes two images (or two directories of frames) to identical resolution/codec, masks HUD regions, COUNTERBALANCES left/right order per trial under a private seed, runs an objective SSIM prefilter, then asks a PINNED vision model which panel is better on each rubric dimension. Returns a win rate with a Wilson 95% confidence interval and a PASS / FAIL / INCONCLUSIVE verdict, plus optional self-calibration that measures the judge's own positional bias and confabulation rate. Turns 'does this look good?' into a measurement.