kind,id,status,bridge,statement,falsify
claim,ts-claim-cf1-contested-claim-level,ready_to_test,climate / NLP / claim verification,"In Climate-FEVER, 19.5% of claims with at least two polar evidence sentences are contested (both SUPPORTS and REFUTES), and that fraction does not rise with the number of polar sentences (k = 2 to 5: 20.6, 16.5, 22.2, 19.3%; trend z = 0.06), whereas independent draws would give 41 to 82%.",A re-annotation or re-retrieval of Climate-FEVER in which the contested fraction among k>=2 claims rises with k (trend z > 1.6) or falls outside 12-28% withdraws this claim.
claim,ts-claim-so1-contested-after-open-retrieval,ready_to_test,CS / NLP / claim verification,"SciFact-Open reuses the 279 SciFact test claims verbatim; none had two polar evidence abstracts in SciFact, and after retrieval over 500K abstracts 81 do, of which 15 (18.5%) are contested, against 71% expected under independence.","If SciFact-Open's claim ids/text do not match SciFact's, or a recount of data/claims.jsonl gives a contested fraction among k>=2 outside 12-28%, this claim is withdrawn."
combination,ts-combo-contested-by-evidence-source,ready_to_test,,"Contestedness is a property of the evidence-gathering process, not of science: retrieval corpora ~20%, replication corpora ~40–60%; the retrieval number measures what annotators find, the replication number measures what experiments find.","A replication corpus re-scored with retrieval-style evidence (papers citing the original) also gives ~20%, or a retrieval corpus restricted to replication studies gives ~40%+."
combination,ts-combo-contested-claims-claim-level,ready_to_test,within-claim polarity concordance: mixed SUPPORTS/REFUTES evidence vs the independence baseline,"In open-retrieval claim-verification corpora the contested fraction among claims with two or more polar evidence documents is about 0.2, independent of the number of documents and of domain; closed citation-built corpora show about 0 by construction, not because scientific claims are uncontested.",A fourth open-retrieval corpus with >=2 polar documents per claim whose contested fraction falls outside 12-28% or whose trend in k has z > 1.6; or SciFact-Open's contested fraction rising above 35% as evidence per claim grows; or independence overprediction below 2x in any such corpus.
combination,ts-combo-listwise-collapse-is-noisy-argmax,ready_to_test,noisy argmax: probability that the best-by-noisy-score is the true best decreases with N at fixed pairwise accuracy,"Accuracy@1 of an LLM pairwise judge over N unexecuted ML candidates equals the noisy-argmax accuracy of a Thurstone case-V comparator at the judge's pairwise accuracy; no additional listwise 'global discrimination' deficit is needed to explain Zheng Table 3, and the same arithmetic bounds RPM child-selection as N grows.","Pre-registered: at p=0.59 the model predicts Acc@1 = 0.221 (N=8), 0.191 (N=10), 0.146 (N=15). A re-run of Zheng's ranking subset at those N with Acc@1 more than 2 SE below these values falsifies the combination; matching values within 2 SE support it. Second test: an RPM/AIRA-dojo tournament with N children whose selection accuracy tracks these curves."