Public Safety & Emergency ServicesIssue 134 records
Emergency Services
Four newly catalogued sources after the 2026-09-17 edition: Sept 16 Daily World report that Grays Harbor Communications deployed Aurelian AVA on non-emergency trunks only (humans remain on 911; 2025 mix ~60% non-emergency) without publishing a local post-go-live diversion audit; an Aug 21, 2026 Nature Communications Earth & Environment perspective proposing governance-constrained multi-agent AI for wildfire interagency coordination while labeling itself conceptual, not operational; NVIDIA/Leidos C2AI multi-agent disaster command-center case study still aimed at operational fielding under human commanders; and Long Beach’s CommsCoach post-call QA pilot with GovCloud processing and 24-hour non-mapped audio deletion as a lower life-safety-risk AI pattern than live emergency triage. Two patterns separate non-emergency automation growth from life-safety proof and bind agentic disaster architectures to explicit human-in-the-loop governance. No new controlled U.S. response-time or clinical-benefit trial dated after 2026-09-17 is claimed.
- Evidence records
- 4
- Cross-source patterns
- 2
- Evidence classes
- 2 independent reporting1 academic research1 vendor claim
- Outcomes
- 3 emerging1 mixed
- Source freshness
- 1 new this fortnight1 recent1 undated1 older, newly relevant
- Research completed
- 2026-09-19
Choose a role to see its takeaway beside every record in the ledger.
Patterns across the evidence
Synthesis · Lighthouse Advisory interpretation
Expand non-emergency automation only with published local KPIs; prefer post-call QA AI when learning ECC controls
Grays Harbor’s Sept 16 AVA launch hard-separates non-emergency AI from human 911 but still leans on vendor 60–75%/90-day averages without a local post-go-live audit in the Daily World story, while Long Beach’s CommsCoach pilot shows a post-call QA path with GovCloud processing, 24-hour non-mapped audio deletion, and human-final scoring. Together they counsel SLED buyers to demand dated local diversion/spillover metrics for non-emergency bots and to treat privacy-constrained QA tooling as a lower life-safety-risk learning path than live emergency triage. This is a control pattern, not proof either city’s tool fails clinically.
Operating questionBefore go-live or renewal, which dated local diversion, spillover, and—when using QA AI—false-flag and retention metrics will be published with human-final decision rules?
Supporting evidenceGrays Harbor Communications (GHCC) / AurelianCity of Long Beach Disaster Preparedness and Emergency Communications / GovWorx
Treat agentic wildfire/disaster C2 as HITL governance work, not fielded autonomy
Nature’s Aug 21 perspective proposes hierarchical multi-agent wildfire coordination but labels itself a conceptual commentary requiring governance alignment and human approval, while NVIDIA/Leidos describe C2AI multi-agent disaster command support still advancing toward operational fielding under human commanders. Sensing and agent frameworks are advancing faster than independent outcome evidence. SLED emergency managers should require observer-mode gates, accept/reject logging, and bans on unsupervised asset tasking before procurement language implies autonomous C2.
Operating questionCan the agency show written HITL policy, observer-mode exercise metrics, and a prohibition on unsupervised fire/EMS tasking before buying agentic C2?
Supporting evidenceUniversity of Cambridge / Caltech / Cal Poly / Open University of Israel / Bard College Berlin (authors)Leidos / NVIDIA
Evidence ledger
Full record · every source keeps its link and limitations
Grays Harbor Communications deploys AVA for non-emergency calls while keeping humans on 911
On 2026-09-16 The Daily World reported GHCC deployed Aurelian’s AVA on the non-emergency line only—explicitly not answering 911—so telecommunicators can prioritize emergency calls and radio traffic; 2025 volumes were 42,014 emergency and 60,596 non-emergency (~60% non-emergency); AVA may create CAD calls for human review, supports 45+ languages, and transfers suspected emergencies; Aurelian claims 60–75% non-emergency handling within 90 days at current agencies—without a published Grays Harbor post-go-live audit in the article.
Why it matters, evidence and limitations
- Why it matters
- Why now: published 2026-09-16 and not archived; newly relevant after 2026-09-16/17 La Vergne and Akron Ava catalogue entries—documents another U.S. consolidated fire/EMS PSAP expanding non-emergency AI with a hard 911 human boundary.
- Evidence and measured results
- Article dated Sept 16, 2026; Green quotes that trained telecommunicators answer 911; 42,014 / 60,596 2025 volumes; nearly 60% non-emergency; CAD create-for-review; 45+ languages; emergency-transfer design; Aurelian CEO quote; vendor 60–75% within 90 days; ~40 agencies / 15M residents / >99.95% uptime since Apr 2024 (vendor).
- Limitations and uncertainty
- Local journalism plus vendor platform averages; no independent GHCC outcome table; combined LE/fire/EMS center—stream focus is fire/EMS capacity and emergency-line protection.
Mitigation of the coordination crisis in wildfire management using a multi-agent AI system
Published 21 August 2026, this Communications Earth & Environment perspective argues that misaligned incentives and fragmented interagency coordination—not only climate and fuels—worsened impacts of the January 2025 California wildfires, and proposes a governance-first hierarchical multi-agent AI architecture with shared global objectives and human-in-the-loop approval of deployment plans; the authors explicitly state the article is a conceptual commentary, not a deployment-ready operational plan.
Why it matters, evidence and limitations
- Why it matters
- Why now: not archived; newly relevant after the 2026-09-17 FireSat first-light catalogue—shifts the wildfire AI conversation from sensing imagery to governance constraints on agentic resource coordination across local/state/federal fire agencies.
- Evidence and measured results
- DOI 10.1038/s43247-026-03962-6; volume 7 article 687; Jan 7, 2025 CA fires cited (31 deaths; $76–131B damages range); NICC/OES mutual-aid hierarchy discussion; ALERTCalifornia >1200 cameras; Technosylva–Pano AI Feb 2026 partnership mention; HITL and gradual observer-role deployment; limitations section labels conceptual commentary.
- Limitations and uncertainty
- Perspective without new field RCT metrics; depends on institutional reforms; international author team with U.S. case focus.
Leidos and NVIDIA build C2AI multi-agent command center for disaster response
NVIDIA’s customer story describes a Leidos–NVIDIA collaboration begun in 2024 on C2AI, a multi-agent architecture that aggregates RF, audio, video, weather, emergency-vehicle locators, and automated 911 transcripts to reduce cognitive load for a human commander, using NIM microservices, Riva Parakeet ASR, Metropolis, and TAK for coordination; the companion Leidos page states agents operate under human supervision and that development continues toward operational fielding—without publishing independent response-time or mortality outcomes.
Why it matters, evidence and limitations
- Why it matters
- Why now: not archived; newly relevant alongside the Nature wildfire agentic governance perspective and recent FireSat sensing updates—documents industry agentic C2 stacks aimed at disaster EM before claiming fielded life-safety results.
- Evidence and measured results
- NVIDIA page: start 2024 collaboration; multi-agent ingest of RF/audio/video; automated 911 transcripts; human commander interaction agent; NIM/Riva Parakeet/Metropolis; TAK mobilization; Unreal Engine synthetic data; agents for video, victims DB, resources, planning, dispatch, commander UI. Leidos companion: human supervision; toward operational fielding; Hendricks quote on 911-to-IC friction.
- Limitations and uncertainty
- Vendor marketing; publicationDate not stated on page; no independent measured outcomes; not claimed as broadly operational.
Long Beach pilots CommsCoach AI to score 911 calls for coaching with privacy guardrails
Hoodline (Apr 29, 2026) reports Long Beach is piloting GovWorx CommsCoach to transcribe and score recorded 911 calls for supervisor coaching (~600,000 annual calls; historic ~2% manual QA; ~$68,000/year through Feb 2028), with privacy materials citing AWS GovCloud processing, 24-hour deletion of non-mapped audio, CJIS/HIPAA framing, and human-final performance decisions—while officials say the AI will not take live calls; some dispatcher skepticism about automated scoring is noted.
Why it matters, evidence and limitations
- Why it matters
- Why now: not previously archived; newly relevant after 2026-09-17 emergency-path and non-emergency PSAP AI catalogue entries—documents a post-call QA control pattern with explicit retention/cloud guardrails as a lower life-safety-risk alternative to live triage AI.
- Evidence and measured results
- Apr 29, 2026 publish; 600k calls/year; ~2% historic review; CommsCoach; $68k/year through Feb 2028; post-call only; AWS GovCloud; non-mapped audio deleted within 24 hours; CJIS/HIPAA claims; supervisor-final evaluations; union/dispatcher skepticism cited.
- Limitations and uncertainty
- Local aggregation; no published accuracy study of scores; April date predates last run but URL was new to archive; policing/fire share ECC—focus is 911 QA controls.
How to read this edition
Source findings, measured results and limitations come from the cited publications. Patterns, operating questions, role takeaways and implementation considerations are Lighthouse Advisory interpretation, stated as questions to validate locally rather than guaranteed outcomes. Vendor and operator claims are labeled as claims. Full research method.
- Independent reporting
- Independent reporting with attributable sources but without a formal evaluation design.
- Academic research
- Research produced through an academic institution or peer-reviewed venue.
- Vendor claim
- A supplier-provided assertion that has not been upgraded to independent evidence.