A hidden reference is the pristine, unimpaired source clip slipped into a test as if it were an ordinary test item, so the viewer rates it without knowing it is the reference. The trick is the heart of the ACR-HR method in ITU-T P.910: you run an ordinary single-stimulus ACR test in which the references are quietly mixed among the processed clips, and the viewer never realises they are doing a comparison. Two things follow. First, it lets you compute a DMOS while keeping ACR's speed - you subtract each viewer's hidden-reference rating from their rating of the matching processed clip (DV = V(PVS) minus V(REF) plus 5). Because each viewer is measured against their own reading of the source, this cancels personal harshness and content preference; a lenient and a harsh viewer recover the same degradation. Second, a viewer's score on the hidden reference reveals their bias. The method only works if the hidden reference is genuinely excellent, since a flawed one distorts every difference computed from it.
