Evidence-controlled disaster-risk software that can ingest live official warning data while refusing to treat a warning as proof of every downstream local consequence.
The system separates official hazard state from downstream consequence, feasibility, responder danger and local action. That boundary is central to the product rather than an afterthought.
The source fabric connects current public-warning and historical hazard evidence to lifecycle/provenance state before any richer consequence or decision claim is permitted to travel.
The evaluation program then tests bounded consequence-priority behavior against explicitly defined comparators rather than claiming to replace official warning operations.
Each item below is deliberately narrower than a product-readiness or superiority claim.
A bounded August 2026 run used the product-owned National Weather Service connector to retrieve current Alabama active alerts and emit new warning lifecycle changes with provenance.
The public-warning path is allowed to carry the official warning state while local consequence/action remains separately evidenced.
The product has source catalog, historical replay and live-source validation surfaces rather than a single demo endpoint.
Earlier AWS evaluation work established the external-judge path with Llama. As current Bedrock custom-metric support changed, the active qualification route moved to Mistral Large. Judge output remains calibration/evaluation evidence and does not override deterministic invariants, live product receipts or confirmatory gates.
It does not claim equivalence to the National Weather Service, warning-issuance authority, autonomous emergency action, confirmed local impact from CAP alone, completed confirmatory superiority, or production public-safety readiness.
The current research/engineering path is closing canonical live-state composition and externally controlled evaluation geometry so consequence-priority claims can be tested without leaking future truth or borrowing authority from the warning source.