# Data Card — SafeEvac Prototype v1

## Primary training data

`data/safeevac_synthetic_training.csv`

- Rows: 6,000
- Origin: synthetic prototype scenarios
- Random seed: 42
- Purpose: hackathon model development and UI demonstration
- Personal data: none
- Individual protected attributes: none

## Included features

### Building / egress
- `floors`
- `exits_total`
- `blocked_exits`
- `stair_width_m`
- `avg_route_m`

### Occupancy and inclusive support planning
- `occupant_count`
- `visitor_pct`
- `mobility_support_count`
- `sensory_cognitive_support_count`

The two support fields are aggregate operational demand fields. They must not be interpreted as a person's value, survival probability or evacuation priority.

### Response organization
- `bhv_count`
- `alarm_delay_s`
- `trained_occupants_pct`
- `communication_channels`
- `night_shift`
- `drill_recent`

### Context
- `sector`
- `hazard`

## Labels

- `risk_index`: synthetic 0–100 operational risk index
- `readiness_score`: 100 - risk_index
- `risk_band`: Low / Medium / High

The label is derived from a transparent prototype rubric plus random variation. It is not an official fire-code or engineering label.

## Empirical grounding used in v1

The generator uses NIST's reported mean stair movement speed of approximately 0.44 m/s as one calibration anchor for route burden. No NIST individual records are bundled into this project.

## Legacy Titanic data

Titanic data is retained only as a bias-learning demonstration in `legacy_titanic/`. It is not merged with SafeEvac operational training data.

## Main limitations

- synthetic scenarios are not a substitute for real drill observations;
- no geometry graph or BIM model yet;
- no dynamic smoke/fire propagation;
- no verified regulatory thresholds;
- model uncertainty is represented only through tree prediction spread;
- no external validation by safety engineers or BHV professionals yet.