Material: served from Ordnance Survey NGD authoritative roof data for ~92% of roofs (badged OS NGD, shown without a hedge). We switched to it after benchmarking our own 10 m spectral guess against it: OS matched 97% of our roofs while the guess agreed only 49% of the time (over-calling glass roughly tenfold). Where OS is absent, a fallback model trained on real OpenStreetMap tags across 12 UK cities, gated on leave-one-city-out generalisation with each city's stock balanced, fills in, and those roofs carry the “Uncertain” hedge plus the top-3 candidates.

Geometry: free 1 m LiDAR (pitch, aspect, height, form) enters as a physics plausibility constraint that re-ranks impossible material/shape combinations (it corrected 4% of materials on a held-out tile), not as a learned feature, which did not generalise to unseen cities.

Albedo: narrowband-to-broadband physics from Sentinel-2 surface reflectance, in-footprint pixels of the least-cloudy recent scene, with a propagated per-roof uncertainty (±0.03 to ±0.07). Estate averages are the robust product; read per-roof albedo with its band. Multi-date stacking and Sentinel-1 / Landsat cross-checks are estate-scale research, not yet in the served per-roof pipeline.

Thermal: peak surface temperature and SRI from the ASTM E1980 balance using reflectance and a per-material emittance, modelled, not measured.

Not yet: in-situ ground-measured roof albedo, measured emittance for the existing roof stock (the coating’s own optics, opaque: 95% reflectance modelled at 0.95 and reference-paint-measured 300-1400 nm, plus 95% thermal emittance, full cured-film characterisation in progress), and per-pixel cloud/shadow masking. We publish uncertainty rather than point claims.