It is a consistency layer, not a veracity layer
What gets measured is agreement between a claim and its evidence, not truth. The distinction sounds like hair-splitting and is the most consequential naming decision in the system.
A layer that scores the relationship between what a model says and what its sources say is often called a truth or veracity layer. That name is wrong, and the wrongness is not academic.
Two sources with overlapping pretraining make the same mistakes together. When they do, they agree perfectly, and any layer measuring agreement reports its highest confidence at exactly the moment the system is most wrong. Truth was never measured. Agreement was.
The name matters because names decide what people build on top. Call it a veracity layer and, eighteen months later, somebody wires an approval workflow to it: claims above 0.9 pass without review. That engineer did nothing unreasonable. They trusted the name.
Call the same number a consistency score and the same engineer asks the next question on their own: consistent with what? That question leads to the source, and the source is where the review belongs.