# SafeEvac Upgrade Log

## What was changed

### Product
- Repositioned from Titanic Survival Predictor to ethical evacuation-readiness decision support.
- Added Readiness Scan, What-if Stress Test, Bias Lab and Data/Governance views.
- Removed person-level survival prediction from the core product workflow.

### AI / model
- Added a new Random Forest Regressor for site/scenario operational risk.
- Added hold-out evaluation.
- Added tree-prediction spread as a transparent model-variation indicator.
- Added counterfactual intervention ranking.
- Added stress testing for blocked exits, communications, alarm delay, night organization and drill recency.

### Data
- Added 6,000 synthetic SafeEvac scenarios.
- Added 17 operational model features.
- Added aggregate support-demand fields for inclusive resource planning.
- Used NIST mean stair movement speed (~0.44 m/s) as one empirical calibration anchor.
- Preserved Titanic separately for bias education only.

### Responsible AI
- Excluded individual protected/socioeconomic attributes from the main schema.
- Added prohibited-field checks.
- Added explicit human-in-the-loop language and deployment boundaries.
- Separated historical bias demonstration from operational recommendations.

### External data review
- NIST building evacuation data: relevant; not bundled yet.
- Snopková et al. route-choice data: relevant, CC BY; not bundled yet.
- Collin et al. 146-drill dataset: relevant/open; recommended for next validation phase.
- Kaggle/WiDS wildfire evacuation-zone dataset: evaluated, not used due use-case mismatch and non-commercial/no-derivatives license constraints.

## What is NOT claimed

- No certified evacuation time prediction.
- No compliance determination.
- No autonomous emergency decision-making.
- No validated life-safety prediction.