Noto Peninsula 2024 replay
The first end-to-end run of the Atlas: five crisis messages modeled on the 1 January 2024 Noto Peninsula earthquake, pushed through the safety, resilience and physics layers.
Setup
- Event: Noto 2024, epicenter area near 37.4°N, 137.1°E (Ishikawa Prefecture).
- Incidents: 5 synthetic messages written to match the kinds of reports seen in the real event.
- Layers: Safety (extract and validate) → Resilience (bundle and route over a simulated mesh) → Physics (wired in, predictions mocked).
- Run date: 5 September 2026.
Results real run, synthetic data
| Incident | Decision | Delivered |
|---|---|---|
| Trapped under collapsed building near Wajima, 3 people | Auto-execute | Yes |
| Water main burst, street flooding, evacuation needed | Auto-execute | Yes |
| Building fire spreading, no water pressure | Auto-execute | Yes |
| Road collapsed, multiple vehicles | Auto-execute | Yes |
| Hospital power out, evacuate ICU patients | Auto-execute | Yes |
{
"earthquake": "noto_2024",
"delivery": { "created": 5, "delivered": 5, "rate": 100.0 },
"validation": {
"latency_under_500ms": true,
"delivery_over_88_percent": true
}
}
Read this honestly
- The 100% delivery rate shows the layers connect. It does not show real-world performance.
- Latency read 0.0 ms because physics inference was mocked. Real latency is not measured yet.
- GPS and hazard checks were hard-coded for this run.
- All five cases auto-executed; no case exercised the human-escalation path yet.
Bugs found along the way real
- Relative-import errors across the safety and resilience modules.
- In-memory SQLite gave each connection its own empty database; nodes now hold one persistent connection.
- NumPy booleans broke JSON output; values are now cast to native types.
Next planned
- Replace synthetic inputs with JMA seismic data and the NIED sensor grid.
- Measure latency with realistic data sizes.
- Compare model output with the damage actually recorded after Noto.
- Add conflicting incidents so the human-escalation path is tested.
- Publish a validation report for domain-expert review.