# Model Card — SafeEvac Random Forest v1

## Intended use

Decision support for evacuation-readiness workshops, audits, training and what-if scenario exploration.

## Prohibited use

- deciding who is saved first;
- denying assistance;
- individual survival scoring;
- autonomous emergency command;
- statutory or engineering certification;
- employment, insurance or access decisions.

## Model

Random Forest Regressor trained on the 17 SafeEvac scenario features.

## Validation

Fixed random 80/20 hold-out split with stratification by risk band. Current expected performance after data generation with seed 42:

- training rows: 4,800
- test rows: 1,200
- MAE: approximately 4.3 risk points
- R²: approximately 0.79

Because the labels are synthetic, these metrics show how well the model learns the prototype scenario rubric; they do **not** prove real-world evacuation accuracy.

## Human oversight

The UI displays recommended interventions as counterfactual scenario changes. A human safety professional must assess feasibility, legal requirements and site-specific context.

## Fairness controls

The primary model schema excludes individual protected/socioeconomic fields. `safeevac_model.py` also includes a prohibited-field guardrail. Aggregate support needs are permitted only for capacity planning.

## Explainability

Instead of person-level feature attribution, SafeEvac ranks operational intervention levers by predicted reduction in scenario risk, e.g. restoring exit availability, improving response capacity, reducing delay, adding communication redundancy and conducting drills.