Numeric and unit contradiction pairs over real documents: same unit differing values, differing units same dimension, overlapping and disjoint ranges, tolerances, precision differences that are NOT contradictions, and conditional qualifications on one side only. Both languages.\n\nCross-language contradiction pairs: the same subject named differently across English and Croatian, contradicting on a value or a decision. These prove the entity work from V2.3 item 6.1 rather than assuming it.\n\nAuthority-ranking cases: which source wins when two disagree, by document class, by revision, or by recency. If that behaviour has not shipped, author the cases and mark them pending rather than gating them, and say so.\n\nAnchoring and fan-out cases as GATES rather than unit tests: the two same-boilerplate datasheets, the multi-model datasheet anchored per section, and the three ambiguity behaviours, all measured in the harness so a future prompt change cannot regress them silently.
Numeric and unit contradiction pairs over real documents: same unit differing values, differing units same dimension, overlapping and disjoint ranges, tolerances, precision differences that are NOT contradictions, and conditional qualifications on one side only. Both languages.\n\nCross-language contradiction pairs: the same subject named differently across English and Croatian, contradicting on a value or a decision. These prove the entity work from V2.3 item 6.1 rather than assuming it.\n\nAuthority-ranking cases: which source wins when two disagree, by document class, by revision, or by recency. If that behaviour has not shipped, author the cases and mark them pending rather than gating them, and say so.\n\nAnchoring and fan-out cases as GATES rather than unit tests: the two same-boilerplate datasheets, the multi-model datasheet anchored per section, and the three ambiguity behaviours, all measured in the harness so a future prompt change cannot regress them silently.