Pramana
Reconstructed replayVerificationProof, a valid means of knowing; how VARUNA scores itself
Where VARUNA is right, where it is wrong, and how we score ourselves.
Scores appear once the event is baked.
Headline scores
Scored from run artifacts against sourced pins.
Contingency by threshold
Pins record waterlogging, not depth, so we score three thresholds.
Rain skill by lead time
Rain skill by lead against the reconstructed truth, with persistence as the bar.
Scoring every baked cycle against the truth field.
Where we are wrong
Pins the model missed, with the deepest water forecast nearby.
Where we were early
Pins flagged before the city logged them.
What we cannot score
Shown rather than omitted.
Method and limitations
Limitations
Limitations we state before anyone asks
From the VARUNA blueprint, section 15.1. Each one is measured on this page as the scores arrive, not hidden behind them.
- Street-level depth accuracy is bounded by terrain data; open DEMs give pattern and timing, LiDAR gives centimetres.
- Radar-based nowcasting cannot foresee convective initiation; beyond ~90 minutes the system is probabilistic by nature.
- Inferred drain networks are hypotheses with uncertainty; the drain-health map is a ranking, not a survey.
- Traffic-as-sensor confounds (accidents, processions) are handled statistically, not perfectly; sensors and CCTV reduce ambiguity.
- Coastal storm-surge coupling uses IMD advisories rather than a full ocean model in V1.