Forecast reliability
Which models remain calibrated during extreme events and geographic transfer?
RESEARCH PROGRAM
Scientific questions before model spectacle.
HeatSafe prioritizes data identity, strong baselines, temporal and spatial validation, uncertainty, sensitivity analysis and reproducible artifacts.
Compound events
When do heat, smoke, PM2.5, humidity and hot nights create non-additive risk?
Urban transfer
Which remote-sensing and morphology features generalize to unseen cities?
Indoor response
Can sparse observations estimate thermal lag and recovery robustly?
AI grounding
Can local and cloud models explain results without altering evidence?
Reproducibility
Which artifacts are necessary for independent replication?
Evaluation commitments
Rolling-origin and chronological splits
Unseen-city and unseen-region tests
Coverage, sharpness and calibration
Missingness and source disagreement