LexRevise-MT v1.2: a benchmark for Legal Conclusion Revision under supervening evidence

LexRevise-MT v1.2 is a benchmark for Legal Conclusion Revision (LCR): labelling, at every turn of a multi-turn legal consultation, how each active conclusion is maintained, revised, withdrawn, or reintroduced as new evidence arrives. It contains 250 four-turn Spanish succession-law scenarios and 4,609 claim-turn labels over four operators (NEW, PERSIST, REVISE, RETRACT), each conclusion anchored to a cited legal basis (norm, article, canonical BOE/ELI identifier, vigency). It provides scenario-level and arc-level splits, a two-round inter-annotator reliability study by three jurists (three-class Fleiss κ = 0.803; a targeted second round raises the persistence/revision boundary to κ = 0.825), deterministic and language-model baselines, and released prediction files. Every reported figure is reproducible from the included scripts (compute_agreement.py, evaluate.py) and fixed by a SHA-256 manifest. Underlying legal texts are referenced by canonical BOE/ELI identifiers rather than redistributed; scenarios are synthetic and pseudonymised. Version 1.2 adopts the second-round gold for the revision stratum and normalises jurisdiction metadata.