Land Regulatory Intelligence (LRI): Benchmark Data and Code

Benchmark data and code for a multi-country vision-language model benchmark on regulatory grounding: first-pass screening of spatial discrepancy candidates between physical conditions observed in remote sensing imagery (de facto) and legal designations in public registers (de jure). Covers Paris, New York City and Tokyo (1,174 tiles) under four input modalities. Includes per-tile raw model responses, Tokyo tile metadata with tile-centre coordinates, the V1 prompt templates, experiment runners, and a script that reproduces the headline tables of the paper. Image rasters are not redistributed.