Quantify recoverable evidence error from PDF-version mismatch

Determine how much of the evidence error caused by PDF-version mismatch is recoverable through improved document acquisition under the final PageRecall configuration.

Background

PageRecall’s evidence evaluation depends on matching the annotator’s PDF revision, because page numbers and evidence locations can differ across document versions. The paper reports that, before the page-selection change, PDF-version mismatch accounted for roughly 32 of 108 development misses. However, the authors did not recompute this proportion for the final system and explicitly state that the recoverable share is unknown. Resolving this issue would clarify how much of the remaining evidence deficit is attributable to acquisition quality rather than retrieval, parsing, or answer-generation errors.