Version-Bounded Filing Question Answering: A Controlled Study of Source Labels, as-of Filtering, and Model Heterogeneity

Abstract

Financial question answering must distinguish a source that contains a relevant disclosure from a source eligible at the question’s cutoff. Evaluation also has to distinguish successful source selection from failures that prevent a narrow error from being counted. We examine these issues in a deterministic synthetic study of 32 original–amendment conflicts crossed with three cutoff regimes. Three prompt interfaces, four attested local model builds, and two fixed repetitions yield 2,304 retained responses. The original internally frozen cross-model hypothesis was unsupported: the mean filtered-minus-mixed wrong-version risk difference was −0.091796875, and only two of eight strata improved. A retrospective descriptive decomposition separates response-contract failure, valid abstention, correct answering, competing-version adoption, and other responses without changing frozen scores. Two builds produced no correct conflict answers in any arm while recording zero competing-version events. Their zeros therefore cannot establish successful version selection. For qwen3:8b, correct conflict answers increased from 48/128 under mixed context to 117/128 under filtering, while counted competing-version events fell from 47 to zero. A worked response shows how a non-Boolean abstention field excludes an otherwise competing answer and citation from that endpoint. Version labels alone gave no consistent benefit. The contribution is an evaluation diagnostic for a finite synthetic suite: response validity, abstention, answer correctness, and source use must be interpreted together. The study does not establish effectiveness on real filings or general temporal reasoning.