Evidence first, when the ground moves.
Pacific Guardian Atlas is an early-stage research platform that pulls Japan's authoritative disaster data into one reproducible evidence lake, and tests how crisis messages can be routed safely when networks fail.
The problem
- Japan publishes world-class disaster data, but it is spread across many agencies, formats and languages.
- Lessons from past quakes live in long committee reports that are hard to compare across events.
- When a big quake cuts power and networks, help requests still have to reach responders.
- AI can read messy reports fast, but it should never be the one deciding who gets dispatched.
What it does
Evidence lake built: v0.1 scaffold
Schemas, ontology and client libraries for 8 official sources: JMA, NIED MOWLAS/Hi-net, J-SHIS, MLIT, Cabinet Office, e-Gov, CiNii/J-STAGE and NIED CKAN/SIP4D. Clients are stubs; live API calls are next.
Zero-trust safety layer built
An LLM may only extract facts from a message. Deterministic code checks them, and conflicts go to a human who picks from pre-validated options.
Offline-first mesh built, simulated
Store-and-forward bundles (RFC 9171 style, Spray-and-Wait routing) and prefecture-level nodes that can answer queries with the central link down.
Physics + cascade models code written, not yet validated
Fourier Neural Operator and spatial-temporal graph network modules for ground motion and infrastructure cascades. Not yet run on real data.
Current status
| Milestone | State |
|---|---|
| Repo structure, database schema, disaster vocabulary | done |
| Safety, resilience and prediction layers wired end to end | done (Sep 2026) |
| Noto 2024 replay with 5 synthetic incidents: 5/5 delivered | done, mock data |
| Real JMA / NIED data for the Noto replay, measured latencies | planned |
| Evidence extraction engine and consensus graph | planned |
| Atlas web UI and GIS hazard tiles | planned |
| Consent-based health-vulnerability layer for evacuation | planned |
Case studies
Noto Peninsula 2024 replay
Five crisis messages through all three layers. What passed, and what was mocked.
Keeping AI out of dispatch
How the zero-trust gate turns an LLM into a reader, not a decider.
Evacuating health-vulnerable people
A planned layer, grounded in the Noto replay's ICU incident.
How to contribute
The project is small and early. These roles would move it furthest:
Data engineers
Turn the stub clients for JMA, NIED, MLIT and Cabinet Office into real, reproducible ingestion with drift checks.
ML / seismology
Validate the FNO and graph models against real Noto ground motion and damage records.
Field partners
Municipal, NGO or hospital teams who can tell us which questions matter in the first 72 hours.
Translators
Japanese ↔ English for policy reports, plus vocabulary for other Pacific languages.
Contact
There is no public inbox yet. Please contact via the Idea Trail portfolio. A GitHub repo and email will be listed here once the project opens up.