FIRELINE

FIRELINE is a Toronto fire-response simulator built for Hack the 6ix 2026. It drops an AI incident commander into building fires across the real city: actual Toronto roads, real fire-station locations, limited apparatus, uncertain reports, and a 3D map where you can watch the response unfold.

It feels more like a strategy game than a chatbot. Alarms arrive, evidence gets stronger or falls apart, engines travel along city streets, ladders and rescue teams take different jobs, and every decision is explained in plain language.

What happens in a mission

1. A seeded fire starts at one or more real Toronto building locations.
2. Callers, alarms, crews, cameras, drones, and social posts send incomplete reports.
3. FIRELINE turns those reports into evidence while blocking prompt injections and invalid data.
4. The commander chooses actions using beliefs and available resources—never hidden simulation truth.
5. Every action is validated before a unit can move.
6. The mission is scored for life safety, stabilization, fire control, speed, and resource use.

The interface includes a fixed high-rise mission, competing incidents, an unsupported industrial hazard, and random Toronto missions with one to eight fires.

Models

• Local safety commander works offline and costs nothing.
• Backboard baseline provides evidence extraction and a generic comparison commander.
• Freesolo SFT is trained on validated Toronto response examples.
• Freesolo final is only promoted if reward training improves held-out results.

Fallbacks are always labelled. FIRELINE never presents a local response as an external or trained-model result.

Run it locally

You need Python 3.10+, Node.js 20+, npm, and a public Mapbox access token.

py -3.12 -m venv .venv
.\.venv\Scripts\python -m pip install -e ".[dev]"
npm install
Copy-Item .env.example .env

Add your keys to .env, then start both services in separate terminals:

NEXT_PUBLIC_MAPBOX_ACCESS_TOKEN=your_public_mapbox_token

.\.venv\Scripts\uvicorn backend.fireline.api:app --reload --port 8000

npm run dev

Open http://localhost:3000. Choose a mission, use Next minute to inspect each decision, or Run mission to watch it play automatically.

Verify it

.\.venv\Scripts\python -m pytest -q
npm run lint
npm run build

Training and evaluation commands are in TRAINING.md. The trust boundary is explained in SECURITY.md and EVIDENCE_ENGINE.md.

FIRELINE is a research simulator, not operational emergency guidance or a real fire-prediction system.