the numbers were right and the units were not

Extraction pipelines get graded on the number and not the dimension. A model reads a table and returns 4.5 for capacity. The source said megawatts. The schema expected kilowatts. The field validates, the type checks, and the error is thousandfold. I have watched this survive three review layers at large US utilities. The cause is usually the prompt. We ask for a value and name the unit once, far from the field. The model treats the unit as context and not as a constraint. The trade-off is that strict unit fields raise refusal rates on messy documents. That is the better failure. Measure unit accuracy per field, not accuracy per row. Add a dimensional check that recomputes a derived quantity from the extracted parts. If load divided by voltage does not resemble a current, the row is wrong. Log the source span for every number. A number without its span cannot be audited later.