Keeping AI out of dispatch
During a disaster, messages arrive as messy text. An LLM is good at reading them. The Atlas lets it read, and nothing more.
The rule
The language model fills in a fixed schema: place, coordinates, type of incident, unverified head count, hazards, urgency and its own confidence. It does no maths, makes no routing choices and cannot dispatch anyone.
How a message flows built
1 · Extract
LLM output is forced into a typed schema with sanity limits (for example, head counts under 10,000).
2 · Check
Deterministic code asks: are the coordinates in Japan? Does the place name match? Does the hazard agree with hazard maps? Is the action threshold met?
3 · Route
All checks pass: auto-execute. Any conflict: an Exception Brief goes to a human operator.
4 · Decide and log
The operator picks from 3–5 pre-validated options. Every decision, automatic or human, is written to an audit trail.
What has been tested
- real The gate ran on all five Noto replay messages, end to end.
- planned Cross-checks against live hazard maps and cell-tower counts. These were hard-coded in the first run.
- planned Adversarial and conflicting messages to exercise the human path.
Why it matters for partners
- Responders keep authority. The model never outranks a person or a sensor.
- Every automated action can be explained after the fact from the audit trail.
- The same gate pattern can wrap any future model, including the planned health layer.