# Production Data Plan

To turn SafeEvac from hackathon prototype into a sellable B2B product, retrain and validate on real evacuation-preparedness evidence.

## Priority 1 — Real evacuation drills

Collect per site/drill:
- timestamp of alarm;
- response/pre-movement delay;
- clearance time by floor/zone;
- route used;
- blocked/unused exits;
- queue/bottleneck observations;
- responder availability;
- assistance events;
- drill outcome and corrective actions.

## Priority 2 — Building and egress data

- floor/zone occupancy limits;
- exit count and status;
- stair/corridor width;
- walking distance to exits;
- floor count/elevation;
- refuge/assembly areas;
- accessible egress provisions;
- optionally BIM/IFC-derived geometry.

## Priority 3 — Inclusive preparedness data

Use operational support categories only:
- number requiring mobility assistance;
- number requiring additional audible/visual/tactile communication;
- number requiring a buddy/escort;
- available evacuation chairs or other approved aids;
- assigned responders.

Do not require diagnoses, ethnicity, income, gender or other identity attributes for the core readiness model.

## Priority 4 — External research datasets

Candidates evaluated for future integration/validation:

1. NIST Building Occupant Evacuation data — 14 buildings, >22,000 measurements; valuable for movement/egress calibration.
2. Snopková et al. route-choice dataset — 208 participants, route choice and decision time in virtual fire evacuation; CC BY and described as usable by academic and commercial entities.
3. Collin et al. experimental evacuation data — 146 evacuation drills, open license; useful for egress-model calibration/validation.
4. 2026 university evacuation drill + Pathfinder outputs — 299 participants and four simulation scenarios; promising for scenario validation, but reuse terms must be confirmed before product integration.

## Kaggle decision

A current Kaggle/WiDS evacuation-related dataset was evaluated but not used. It predicts wildfire time-to-threat for evacuation zones, which is a different problem, and its stated CC BY-NC-ND license is a poor fit for a commercial SafeEvac core model.

## Pilot recommendation

Start with 3–5 buildings from one partner (e.g. BHV provider, care organization or facility/security provider), run controlled drills and build a site-level dataset before claiming operational predictive validity.