A benchmark built with Ellis Solaiman, published at ICNLSP 2026.

Given a short piece of civic evidence and a claim, does the evidence support it, refute it, or not say? Real civic text makes that harder than it sounds. Two sentences can be nearly identical and still disagree, because one quietly changes a place, a time, a status or a cause. CURB targets exactly those cases, across transport, roadworks, weather and crime.

LaSSI reaches 85.6% accuracy and 0.858 Macro-F1, narrowly ahead of the best natural language inference cross-encoders, with its clearest wins on location, time and cause. Paraphrase and status changes still favour the neural models, and terse roadworks notices are where the parser struggles most.