Adjudicated human verdicts on atomic visual claims in machine-generated image descriptions

Detailed descriptions of photographs were generated by three vision-languagemodels (Qwen2.5-VL-7B-Instruct, Idefics3-8B-Llama3, LLaVA-NeXT Mistral-7B),decomposed into atomic visual claims, and each claim assigned one of threeverdicts: supported (the photograph shows this to be the case), contradicted(the photograph shows this not to be the case), or unverifiable (the photographcannot settle the question either way). A stratified sample of 501 claims was labelled independently by two annotators,who disagreed on 87. All 87 disagreements were adjudicated in a single jointsession, recorded here claim by claim: the verdict each annotator held going in,the verdict that became the consensus, which annotator changed position, and theagreed cause of the original disagreement. Twelve claims reached no agreedverdict and are released as such rather than discarded. This release contains the annotation protocol in the version frozen beforedouble annotation, the pre-registration dated before the adjudication session,both annotators' independent verdicts with uncertainty flags, free-text reasonsand region markers showing where in each photograph the evidence was found, thefull adjudication record with cause codes, provenance flags distinguishingspontaneously agreed from adjudicated labels, and verifier confidence scores onboth the agreement-filtered and adjudication-expanded gold sets. Images are drawn from the intersection of COCO 2017 and Visual Genome and arenot redistributed here; image_id is the COCO image id. See README.md for thecolumn dictionary and for caveats on gold-set construction, label provenanceand claim-type reliability.