Every number describes a bounded relative index and a dated evidence snapshot—not a prediction of water depth.
S = 0.60 H + 0.25 D + 0.15 U (relative susceptibility); F is bounded recent/forecast precipitation; E decays approved positive observations. The resulting R = clip(0.05 + 0.65 SF + 0.15 F + 0.60 E, 0, 1).
Weather comes from a model grid, not a road-level water sensor. The score cannot be read as a probability that a particular road will flood.
Evidence coverage combines weather freshness, terrain coverage, graph age and local observations. If precipitation windows are missing or older than 60 minutes, comparable risk scores are withheld; active reviewed closures still apply.
Candidate searches are bounded. Routes must obey road direction, turns and reviewed closures; displayed lower-risk options fit within your chosen 0–30% estimated-time detour. The candidate-set best is not a globally optimal route or a promise of safety.
Source registry: OpenStreetMap / ODbL 1.0, Open-Meteo / CC BY 4.0 and API terms, Copernicus DEM (artifact license review required), Delhi Traffic Police official publications, and an optional Apache-2.0 tagged pretrained classifier. See project overview.
Source-derived, observed and synthetic artifacts are separately labeled. No independent model accuracy benchmark is claimed.