A Frozen Foundation Model Is No Better Calibrated Than Six Spectral Indices: Auditing Urban Heat Maps Across Cities

Cities decide where to plant shade and lay cool pavement using machine-learned heat maps, increasingly built on geospatial foundation models that a municipality consumes frozen and cannot retrain. Those maps carry error bars, and whether the error bars hold in the neighbourhoods where the money lands has not been checked. We check, in four cities on three continents (Phoenix, Seattle, Chicago, Karachi), using public Landsat and Sentinel-2 imagery and the global Local Climate Zone map. A prediction interval advertised at 90 percent is valid city-wide everywhere we look (0.888 to 0.901) while its coverage across urban zones spans 0.736 to 0.978, and falls to 0.105 when a model calibrated in one city is applied to another. More surprisingly, a frozen 1024-dimensional Clay v1.5 embedding is no better calibrated under that transfer than six hand-built spectral indices: mean coverage 0.513 against 0.480. A recent benchmark found foundation models are no more accurate out-of-distribution than simple baselines; we find the same is true of their uncertainty, which is the property a planner actually relies on. We argue the remedy must sit outside the model, sketch a label-free conformal repair, and release an audit that runs on a laptop from public data.