Put a minimum count in the JSON schema you paste into the prompt — never enforced, just stated — and the share of extraction lists longer than the document’s true count goes from 34.5% to 55.4%. Enforce it in the decoder and it reaches 90.1%. The extra entries are mostly real text lifted out of the document, so the guard you already have — check each entry appears in the source — passes most of them. Across 8 open-weight models, 558 real documents and 9 annotated sets, a shown minimum adds 0.62 entries on average, and re-asking with the validation error makes it far worse.