Code and data for "Frontier AI agents know where the world's people live but not what their cities emit"

Code and data accompanying Song, Ma, Ran & Gao, "Frontier AI agents know where the world's people live but not what their cities emit". The deposit contains the 11,413-city truth table and stratified 400-city sample derived from the GHS Urban Centre Database R2024A, the city records given to four frontier AI agents under three conditions (named, coordinates-only, anchored), the session instructions and raw outputs of 52 audit sessions and 8 replication sessions, location-only reference-model predictions, the 5,200 scored records, re-identification adjudication, all analysis and figure code (Python), and the Source Data (Data S1). See README.md for layout and reproduction steps.