Public Safety & Emergency ServicesIssue 624 records
Emergency Services
Four signals on keeping emergency AI in an evaluation box before it changes live operations. A 3 November Victoria University of Wellington comment says an MBIE-funded dispatch model, replayed by Omda on one year of de-identified Wellington Free Ambulance history, improved average and 90th-percentile response times across all urgency categories; a six-month passive parallel run is planned for mid-2027, and the page publishes no seconds-saved figure. Brazil’s science ministry reported on 6 November that Cemaden’s FloodcastingXAI used 10-minute rain and river-gauge data to nowcast flash-flood rise one to four hours ahead, with operational tests showing about 40% better forecast performance—without naming the metric—and with humans still issuing alerts. Aspen Public Radio (6 November) describes Aspen Fire’s donor-funded plan to operate Seneca AI-supported suppression drones as early as next year after prescribed-burn tests and a late-October FAA waiver limited to 200 feet and visual line of sight; an independent expert says the aircraft may be too small and have not fought real wildfires. NIST PSCR’s 5 November write-up of the INSIGHT indoor-mapping pipeline reports that older 3D models scored near zero on safety fixtures and requires a human reviewer because unverified maps could endanger crews.
- Evidence records
- 4
- Cross-source patterns
- 2
- Evidence classes
- 2 government evaluation1 academic research1 independent reporting
- Outcomes
- 4 emerging
- Source freshness
- 4 new this fortnight
- Research completed
- 2026-11-08
Choose a role to see its takeaway beside every record in the ledger.
Patterns across the evidence
Synthesis · Lighthouse Advisory interpretation
Emergency AI is still being held to historic replay, prescribed-burn tests, or human-reviewed pre-incident use before live control
Wellington’s dispatch model is credited with better average and tail response times only on a year of historic ambulance data, with a mid-2027 passive parallel run before any live effect. Aspen Fire’s AI-supported suppression drones have been tested on prescribed burns, not wildfires, and may fly real fires only under a tight FAA waiver. NIST confines INSIGHT to pre-incident mapping because unverified indoor products could endanger crews. Agencies should treat these as evaluation designs, not proof that AI is ready to assign ambulances, drop retardant, or navigate interiors on its own.
Operating questionWhat independent historic or constrained-field evaluation must be finished before an AI system is allowed to change live EMS assignment, aerial suppression, or interior navigation?
Supporting evidenceVictoria University of Wellington / Weewoo / Wellington Free AmbulanceAspen Fire Protection District / SenecaNIST Public Safety Communications Research (PSCR) Innovation Lab
Dispatchers, flood monitors, firefighters, and map reviewers remain the stated decision authority around new emergency AI
The Wellington model presents recommendations that dispatchers can accept, change, or ignore. FloodcastingXAI is described as adding explained nowcasts for Cemaden and civil-defense staff, not replacing alert issuers. Aspen crews still command the swarm under continuous visual observation. NIST requires a human to correct the generated indoor graph. The operating question is whether that human authority is staffed, logged, and audited—or only asserted.
Operating questionHow will an agency document, staff, and audit the human accept/override so it is real when AI recommendations arrive in seconds?
Supporting evidenceVictoria University of Wellington / Weewoo / Wellington Free AmbulanceCemaden / Ministério da Ciência, Tecnologia e Inovação (MCTI)Aspen Fire Protection District / SenecaNIST Public Safety Communications Research (PSCR) Innovation Lab
Evidence ledger
Full record · every source keeps its link and limitations
Wellington ambulance ML recommendations improved simulated average and 90th-percentile response times; live parallel run slated for mid-2027
A 3 November 2026 Victoria University of Wellington comment by PhD researcher Jordan MacLachlan describes a machine-learning sidecar, funded by an MBIE Smart Ideas grant and supervised by Professor Yi Mei, that presents dispatch recommendations in a fraction of a second which dispatchers can accept, change, or ignore. Identical situations return identical recommendations so decisions can be audited. In late 2024, Health New Zealand asked publicly traded Norwegian firm Omda to simulate one year of de-identified historic Wellington Free Ambulance emergencies; compared with a dispatcher-style benchmark co-designed with the service, the model improved average and 90th-percentile response times across all urgency categories. That result led Health New Zealand, ACC, and Wellington Free Ambulance to support a phased real-world trial. A six-month passive parallel run on data de-identified in real time is planned to start in mid-2027, comparing every dispatcher decision with the model without affecting live dispatch. The opened page does not publish a seconds-saved figure. MacLachlan’s team has started a company, Weewoo (six people). Context cited: Hato Hone St John and Wellington Free Ambulance attended more than 640,000 incidents in the 12 months to July 2026, up more than 12% on the 12 months to July 2022; the average Wellington paramedic lasts only seven years on the road, with roughly 28% still working a decade after registration versus around 73% of doctors.
Why it matters, evidence and limitations
- Why it matters
- Relevant to U.S. EMS medical directors and 911/ECC boards considering AI that recommends which ambulance goes where: this is third-party historic simulation plus a planned passive parallel run, not a live deployment. Transferability is limited by New Zealand fleet rules, urban/rural mix, and a researcher-authored account from the founder of Weewoo. Why now: published 3 November 2026 and confirmed not in the emergency-services archive, and it is the first public description of the Health NZ/Omda year-long replay and the mid-2027 Wellington Free Ambulance parallel evaluation.
- Evidence and measured results
- University page dated 3 November 2026 (article:published_time 2026-11-03T16:42:41+13:00). First-person account of Omda’s historic-year simulation and partner support for a mid-2027 six-month passive parallel run. No published effect-size in seconds on this page.
- Limitations and uncertainty
- First-person commentary by the researcher who founded Weewoo; Omda results are summarized without a published seconds-saved figure, confidence interval, or open methods paper on this page; simulation of historic year is not live, messy operations; New Zealand context does not automatically transfer to U.S. PSAPs.
Brazil’s Cemaden says FloodcastingXAI operational tests improved 1-to-4-hour flood forecasts by about 40%
On 6 November 2026 (18h16), Brazil’s science ministry reported that Cemaden researchers developed FloodcastingXAI, an explainable-AI platform that ingests rain-gauge and river-level observations updated every 10 minutes to forecast water-level rise one to four hours ahead for civil-defense decision support. The ministry says the system does not replace monitoring staff or alert issuers; it adds information to technical analysis and can show which inputs most influenced a forecast. Operational tests reportedly showed gains of approximately 40% in forecast performance for one- to four-hour lead times; the release does not name the skill score or baseline behind that figure. FloodcastingXAI is a product of the iFast project (more than 60 researchers at 15 Brazilian and foreign institutions) with collaborators including Inpe, Unesp, Unifesp, and the City of São Paulo. Development used open source without external contracting. The tool is a finalist in the 30th Enap Public Sector Innovation contest (AI for the Public Good category; more than 700 entries; competing with five other projects in that category; popular vote open until 11 November; initiative no. 983), led by Leonardo Bacelar with Cemaden teammates named in the release. It previously received an AI4GOOD prize at Brazil Conference 2025. The same article mentions Cemaden’s separate GeoRisk landslide tool (up to 72-hour risk forecasts) as a related but distinct system.
Why it matters, evidence and limitations
- Why it matters
- Transferable to U.S. emergency-management and flood-prone local governments that already operate gauge networks and need short-fuse nowcasts for evacuation or road-closure decisions. Limits: Brazilian hydrology, sensor density, and an unspecified 40% performance metric in a ministry news release rather than a peer-reviewed skill-score table. Why now: published 6 November 2026 and confirmed not in the archive.
- Evidence and measured results
- MCTI news page stamped Publicado em 06/11/2026 18h16. Direct statements on 10-minute gauge updates, 1–4 hour lead times, explainability, non-replacement of alert staff, approximately 40% operational-test gains, iFast partnership counts, open-source development, and Enap finalist status.
- Limitations and uncertainty
- Ministry self-report of its research unit’s tests; the 40% figure has no named metric, geography, or independent audit in this article; hydrologic transfer to U.S. basins is unproven here; Enap finalist status is not a performance evaluation.
Aspen Fire plans 2027 launch of AI-supported Seneca suppression drones after prescribed-burn tests and a limited FAA waiver
Aspen Public Radio reported on 6 November 2026 (9:02 p.m. MST) that the Aspen Fire Protection District is training to become, once staff finish training, the first U.S. fire department—and, the story says, the first in the world—to operate Seneca’s AI-supported firefighting drones independently. Reporter Halle Zander observed 24 August 2026 training in unincorporated Pitkin County. Each aircraft is described as about 10 by 12 feet, able to fly up to 20 miles empty, and designed to work in a swarm; together, five drones can hold 500 to 1,000 gallons of finished fire foam. When they arrive at a fire they use AI to calculate distance and wind speed, then release water or retardant at a specific angle and velocity. Volunteer firefighter John Hammarback called the system “pretty self-flying,” with obstacle stopping; Deputy Chief Andy Fisher wants them sent ahead to remote lightning starts. Seneca founder Stuart Landesberg compared them to a “flying Waymo,” said they will not replace air tankers on large fires, and claimed they can be “the next best thing at two orders of magnitude less cost.” Aspen Fire says it will pay about $500,000 a year over five years (about $2.5 million), funded by private donors. Charles Werner of DRONERESPONDERS noted the drones have only been tested on prescribed burns, not real wildfires, and said he does not think their size will be sufficient for what the department wants to accomplish. Aspen Fire secured an FAA waiver in late October and hopes to use the aircraft on real fires as early as next year, limited to 200 feet in Pitkin County, not over populated areas, and only with eyes on the aircraft at all times—no sending a drone blindly over a ridgeline.
Why it matters, evidence and limitations
- Why it matters
- Direct U.S. fire-service signal for departments watching autonomous initial attack: cost, donor funding, FAA visual-line-of-sight and altitude limits, and an independent expert’s size/sufficiency caution. Related Seneca partner announcements (including Pano AI and a San Bernardino County contract) are already in the archive; this is a different URL focused on Aspen’s operational constraints rather than vendor launch copy.
- Evidence and measured results
- Halle Zander report dated 6 November 2026 at 9:02 p.m. MST (page metadata 2026-11-07T04:02:00Z) with on-scene 24 August 2026 training, named firefighter and vendor quotes, FAA waiver limits, and Werner’s prescribed-burn and size caveats.
- Limitations and uncertainty
- No measured wildfire containment results; Seneca and department claims about ease of use and cost are not independently audited here; Werner explicitly questions whether current airframes are large enough; “first in the world” is the reporter’s/department framing; related vendor stories already archived should not be treated as new performance evidence.
NIST PSCR’s INSIGHT pipeline uses foundation models plus a required human reviewer to turn building scans into first-responder indoor maps
In a 5 November 2026 article, NIST’s PSCR Innovation Lab describes INSIGHT (Indoor Scene Intelligence from Geometric-Semantic Hierarchy Transfer), an automated pipeline that converts building scans into structured data for pre-incident planning, with later phases envisioned for live tracking and command. PSCR argues indoor disasters lack the satellite, GIS, GPS, and drone coverage available outdoors, while 2D blueprints often omit standpipes, shutoffs, alarm panels, and AEDs. NFPA 1660 and NFPA 950 require detailed, interoperable pre-incident data that departments now produce manually. Purpose-built 3D models “score near zero” on small, visually varied safety fixtures (exit signs, alarm panels, AEDs); PSCR says it measured that gap in a 2025 cross-dataset analysis using Point Cloud City, and that foundation models can recognize those features without task-specific training, with registered depth placing them in 3D. The pipeline is designed to follow the NIST AI Risk Management Framework by keeping a human reviewer in the loop to correct the generated graph, using independent detection paths that flag disagreement, and applying building-code bounds to implausible results—because “without proper verification, results could endanger first responders.” Outputs are intended to reach responders in a matter of seconds over existing broadband, use public data and open-source tools, follow ISO 19164, IndoorGML, and IFC naming rather than a proprietary format, and support role-filtered views for dispatch, incident command, and discipline-specific users. Current work is strictly pre-incident.
Why it matters, evidence and limitations
- Why it matters
- Directly applicable to U.S. fire departments struggling to meet NFPA pre-incident data expectations at scale. The piece is as much a governance and safety-control note (human review, dual paths, AI RMF) as an accuracy claim. Why now: published 5 November 2026, confirmed not in the archive, and missed by the 2026-11-06 edition; it is PSCR’s public description of INSIGHT and the 2025 near-zero fixture-detection measurement.
- Evidence and measured results
- NIST page dated November 5, 2026 (time datetime 2026-11-05T12:00:00Z). Describes INSIGHT, the 2025 cross-dataset finding, AI RMF human-review design, NFPA 1660/950 context, open standards, and the explicit endangerment warning if outputs are unverified.
- Limitations and uncertainty
- This is a NIST lab news article about a pipeline under test, not a published accuracy table for INSIGHT itself; “near zero” refers to purpose-built 3D models on safety-fixture classes in a 2025 analysis, not a full INSIGHT field trial; live tracking remains future work; no timed comparison versus manual pre-plans is given beyond qualitative “seconds” delivery and “exponentially less work” for reviewers.
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.
- Government evaluation
- A public body’s measured evaluation or documented pilot.
- Academic research
- Research produced through an academic institution or peer-reviewed venue.
- Independent reporting
- Independent reporting with attributable sources but without a formal evaluation design.