FACTUAL-SUPPORT REVIEW — HUMAN PROCESS, NOT AN AUTOMATIC ENTAILMENT MODEL
1 Split the proposed answer into atomic claims. Compound promises may require separate labels.
2 Check citation eligibility: correct document ID/version/status and permitted user scope.
3 Locate an exact quote; read enough surrounding context to understand scope and conditions.
4 Compare entity, time, number, unit, negation, modality ('may' versus 'will'), quantifiers
('some' versus 'all') and required conditions. Similar wording alone is insufficient.
5 Label supported ONLY when current eligible evidence establishes the whole claim.
Label contradicted for an explicit conflicting assertion. Label insufficient for absent,
outdated or incomplete evidence. Missing quote does not by itself prove contradiction.
6 Record claim, source/version, quote, label, reason and uncertainty; request human escalation.
7 For a real evaluation use independent reviewers and resolve disagreement with evidence.
These 8 authored cases illustrate the rubric; they do not validate automatic semantic judgment.
Exact quote membership can pass for a false paraphrase. Embedding similarity is not entailment.
No job, ATS, employment, factual accuracy or universal safety guarantee is implied.
