Evidence-governed decision support for disaster operations. The system ingests live and historical hazard data, preserves source timing and provenance, and keeps hazard, consequence, response feasibility, responder danger and action authority separate before producing bounded action candidates, holds, source demands and decision receipts.
An official warning can establish that a hazard exists without proving what happened at a specific property, whether a route is passable, whether response is feasible, whether responders are exposed, or what consequential action is authorized.
The system uses typed source adapters and canonical incident/decision state to preserve authoritative hazard evidence separately from downstream consequence, response feasibility, responder danger and local action authority. Claim ceilings and fail-closed gates prevent one source from silently carrying conclusions it does not support.
Live and historical source pathways can be replayed, fault-injected and compared under preserved provenance so model-assisted reasoning and evaluator outputs remain subordinate to deterministic state, source timing and action permissions.
Each item below is deliberately narrower than a product-readiness or superiority claim.
A live AWS runtime used the product-owned National Weather Service connector to ingest a current warning and serialize canonical incident, surface, decision and state objects with source timing and provenance preserved.
Current source work includes live and historical hazard pathways, temporal replay, source lifecycle state and fault injection so warning, event history and downstream decision evidence can be tested without collapsing them into one signal.
Validation has exercised Dockerized runtimes, GitHub Actions, PostgreSQL 17, local Supabase and S3 versioned artifact custody so incident and decision state can be tested with persistent evidence and reproducible execution surfaces.
Controlled evaluator work tests whether hazard evidence remains separated from consequence, response feasibility, responder danger and action authority. AWS Bedrock has been used for model-evaluation workflows, including NVIDIA Nemotron 3 Super 120B, while deterministic invariants and live product receipts remain controlling evidence.
It does not claim equivalence to official warning authorities, warning-issuance authority, autonomous emergency action, confirmed local consequence from a warning alone, field/public-safety efficacy, completed confirmatory superiority, guaranteed loss reduction, or production command authority.
The executed product body includes live official-source ingestion, typed hazard/decision state, provenance, replay, fault injection, database/container/cloud custody and a controlled evaluator program. Selectively scoped pilots remain possible where an organization brings a defined hazard, consequence or response workflow whose evidence boundaries and success criteria can be explicitly bounded.