A Multi-Axis Pre-Deployment Predictor of Cross-City Fairness-Audit Transfer in Spatial Crime Models: Evidence from a Twenty-Two-City UK Cohort

We ask whether failure of a cross-city fairness-audit transfer in spatial crime models can be predicted before deployment from properties of a holdout city, where audit failure is the joint criterion of (mean |ΔDIR|, max |ΔDIR|, count of |ΔDIR| > 10) and DIR is the Disparity Index Ratio of top-decile predicted-violent-crime flagging in above-median versus below-median demographic cells on a given axis. We assemble a twenty-two-city UK production cohort (nineteen Tier 2 cities plus three Tier 1 baselines) using police incident data from data.police.uk over a 35–36 month window, fit per-city violent-crime spatial models on H3 r9 hexagonal cells (~0.105 km²) using a gradient-boosting ensemble, and audit DIR on three minority axes from the 2021 UK Census and one deprivation axis from the English IMD 2019 and Welsh WIMD 2019. We construct four single-feature pre-deployment predictors, calibrate class boundaries on a 22-city retroactive set, and test two post-calibration cities (Southampton, Portsmouth) prospectively. The four-axis aggregate score yields perfect boundary separation of the Tier 2 cohort: score = 0 maps to audit PASS for 4/4 cities; score ≥ 2 maps to non-PASS for 10/10 cities. A six-mechanism failure taxonomy assigns every non-PASS verdict to a predictor-detectable mechanism family. Event-probability calibration via Brier score with Murphy decomposition across 46 (city, type) pairs shows all 46 calibration slopes within the conventional 30% deviation flag threshold. Within the UK jurisdictional scope, pre-deployment city properties are sufficient to predict cross-city fairness-audit failure; findings should not be generalized cross-jurisdictionally without recalibration.