Captured from the live deployment, not a mockup
01 · The problem
Modelled Bengaluru road-capacity loss over 298k violations on one H3-by-hourly substrate, with strictly causal lag features at t-1h, t-24h and t-168h so nothing leaks from the future.
02 · How it works
- 01
Engineered spatial features as H3 at three resolutions with k-ring neighbour aggregates and target-encoded stations, plus OSM built-environment signals including junction degree and edge betweenness.
- 02
Benchmarked 8 model families against a 0.291% label rate, reported a 46.9x lift over base rate, and flagged the original PR-AUC target as unreachable by construction rather than presenting the shortfall as a weak result.
- 03
Built one H3-by-hourly substrate over 298k violations and 8k events, then benchmarked 8 model families on it with strictly causal lag features so nothing leaks from the future.
- 04
Published an enforcement-bias audit of the training data itself, and reported one task as not-yet-answerable after finding the review process changed regime mid-window, with rejection jumping to 0.790 from about 0.28.
03 · What it cost, and what it returned
Caught a zero-hotspot Gi* result as an artifact of the permutation p-floor rather than a finding, and documented it alongside the other published failures.