Earthquake Early Warning

Frontier If seconds of warning actually convert into fewer injuries and less broken equipment — which no study has measured, and section 7 says exactly what would settle it — the prize is a continent-scale reflex that fires in under ten seconds and costs roughly one highway interchange a year. The condition goes first because the engineering is finished and the evidence is not. ShakeAlert has been issuing public alerts in California since 17 October 2019, in Oregon since March 2021 and in Washington since May 2021; Japan has run a nationwide system since 2007 and Mexico since 1991. What none of them has produced is an outcome number.

Established The magnitude available, stated in the only unit that matters, is seconds, and the measured values are small and precisely known. ShakeAlert’s fastest first alerts arrive 4–8 s after earthquake origin time in well-instrumented areas; during the 2022 M 6.4 Ferndale earthquake, half of the sites that went on to record Modified Mercalli intensity VIII shaking received their alert more than 10 s before that shaking arrived. Across 30 simulated M 9 Cascadia ruptures, just under 60% of the Pacific Northwest population would be alerted before MMI V shaking reached them. The baseline against which all of that is measured is zero: before these systems, everybody found out when the floor moved.

Established The cost side is small and fully public, which makes this one of the few hazard technologies where the appraisal is bounded on one side. Congress appropriated $161.8 million for United States earthquake early warning between FY2006 and FY2024, rising to $28.6 million in the final year, for a network that stood at 1,509 stations in December 2023 against a target of 1,675. Set that against the $2–3 billion attributed to injuries alone in the 1994 Northridge earthquake, and the arithmetic looks easy — but the arithmetic is the easy half, and section 3 shows the hard half has never been done.

Speculative The upside is also distributed badly, and the distribution is set by physics rather than by policy. The blind zone — roughly 25 km of radius around the epicentre at current latencies — is where shaking is worst and where nobody can be warned. Mexico City’s famous 50 s of warning is a gift of geometry, not of engineering: it comes from sitting several hundred kilometres from the subduction zone, and it vanished in September 2017 when the rupture happened underneath. The condition on everything above is a single missing measurement: whether a facility wired to act on an alert is measurably safer than an identical facility that is not, which section 5 names as the binding constraint on the whole case.

Established Earthquake early warning is not prediction; it is a race between two signals from an earthquake that has already started. A rupture radiates compressional P waves that travel at roughly 6 km/s through crustal rock and shear S waves plus surface waves that travel at roughly half that speed and carry most of the damaging energy. A seismometer near the source detects the P wave; a processing centre estimates location and magnitude; a message travels at the speed of light down fibre and radio. If the message wins the race against the S wave at your location, you get a warning. If it does not, you do not. Nothing about this requires knowing that an earthquake will happen; it requires knowing, very fast, that one has.

Established The warning time at any point is a difference of two travel times minus a fixed system latency, and that identity is the whole subject. The United States Geological Survey’s own accounting puts communication delays through public channels at 2–3 s and average user reaction time at a further 5 s, which together define a no-warning radius of about 25 km around the epicentre. Denser instrumentation close to a fault can compress that: a 2022 review by Festa and colleagues at Naples puts the achievable floor at 15–25 km for networks specifically built around a known fault zone. It cannot be driven to zero, because the detection itself requires waves to reach at least one instrument.

Established Two architectures compete, and they fail in opposite directions. Regional or network-based systems locate the source, estimate a magnitude, and predict shaking everywhere from a ground-motion model; they give longer warnings at distance and are vulnerable to source-parameter error. On-site systems measure the first arrivals at the protected structure itself and trigger on a local threshold; they are fast and cannot be fooled about magnitude, but they give almost no lead time. The Chinese national system, the Japanese system and ShakeAlert are all now hybrids, running several algorithms in parallel and taking the maximum predicted intensity.

Established The third architecture is wavefield propagation, which sidesteps magnitude entirely. The Japan Meteorological Agency’s PLUM method — Propagation of Local Undamped Motion, in operation since 22 March 2018 — predicts shaking at a site from shaking already observed at nearby sites, without estimating a magnitude or a hypocentre at all. It cannot give long warnings, because it waits for real shaking somewhere, but it does not saturate and it does not care whether the rupture is a point or 500 km of fault. This is the single most important design idea to come out of the 2011 Tohoku failure.

Established Magnitude saturation is the specific physical reason large earthquakes are hardest. An algorithm that estimates magnitude from the first few seconds of the P wave is reading the beginning of a rupture that may continue for two more minutes. Minson and colleagues put the constraint formally in 2018: strong ground motion cannot be expected at a site unless the earthquake is either very close or has already grown very large, and since magnitude is not predictable, the time before strong shaking can be declared is bounded below. Their counterintuitive corollary — that users who ask to be alerted at lower shaking thresholds get more robust warnings with longer average lead times than users who ask only for severe shaking — is the design rule every operational system has since had to confront.

Established ShakeAlert’s first four years of public operation are documented to a level of detail unusual for any deployed warning system, and the headline is that detection works and edge cases do not. Lux and 21 co-authors reported in the Bulletin of the Seismological Society of America in 2024 that between 17 October 2019 and 1 September 2023 the system created 95 events with maximum magnitude estimates at or above the M 4.5 public alerting threshold, of which 94 were real earthquakes. Seven were categorised false under the system’s internal definition — no matching catalogue event within 100 km and 30 s of origin time — but all but one of those seven were real earthquakes badly located at the edges of the network. Three further detected events were labelled missed because their location error exceeded 100 km, and the system produced no solution at all for four catalogue events of M 4.5 or greater, every one of them at the boundary of the network.

Established Read as a miss rate against the catalogue rather than as a detection count, the same record fails the system’s own standard. The Congressional Research Service’s April 2025 report puts it flatly: ShakeAlert sent alerts for 41 of 53 qualifying earthquakes, with seven mislocated and five having magnitudes underestimated below the alert threshold, and “the miss rate of 22.6% is above the performance standard of a 10% or less miss rate set for ShakeAlert.” Both statements are true and they come from the same underlying study. Which one a reader meets first determines whether the system reads as a success or a shortfall, and this brief prints both because the disagreement is about framing, not about data.

Established Latency is the part of the system that has been engineered hardest and is now close to its floor. Fastest first alerts run 4–8 s after origin time in well-instrumented areas. Specific events: the 2021 M 5.3 Calipatria earthquake produced its first alert 6.7 s after origin with a 1 km location error and an initial magnitude estimate of M 5.2 against a final M 5.3; the 2022 M 5.1 Alum Rock earthquake alerted at 5.0 s; the 2022 M 6.4 Ferndale earthquake alerted at 7.5 s. Delivery to phones adds its own tail: for Alum Rock, 20% of MyShake handsets reported latencies under 2.2 s, 50% under 3.5 s and 80% under 6.0 s. The slowest fifth of the delivery distribution is comparable in size to the entire detection budget.

Established Warning times actually delivered, at the shaking levels that matter, are short and highly variable. For the 2021 M 6.2 Petrolia earthquake, sites that recorded MMI VI got 11–23 s of warning using the MMI III alert contour. For Ferndale in 2022 the equivalent range at MMI VI sites was 0 to 23 s — the lower bound is not a typographical flourish, it means some people who felt strong shaking received the alert as the shaking began or after. At MMI VIII sites in the same event, half got more than 10 s. This is the real shape of the product: excellent at the fringe of the felt area, marginal in the damage zone.

Established Magnitude error is where the system is worst, and the failures are instructive rather than random. Ferndale peaked at an estimated M 6.7 against a final M 6.3 — an over-alert. The Truckee earthquake, finally M 4.7, was estimated at M 6.0 with a 29 km location error 17.8 s in, then corrected downward to M 3.7. A single event thus produced a two-and-a-half-magnitude-unit swing in the operator’s own estimate inside twenty seconds. Wireless Emergency Alerts have been issued for 21 earthquakes, ten of which had catalogue magnitudes below 5.0, which is the price of a threshold set deliberately low.

Established The build-out is nearly finished and its cost is public. The network held 1,509 broadband and strong-motion stations in December 2023 against a target of 1,675 that the operators expected to reach by 2025; the Congressional Research Service put it at 1,553 stations and 90% complete in December 2024. Federal appropriations totalled $161.8 million from FY2006 to FY2024, with the annual figure rising from $6.5 million to $28.6 million over FY2015–FY2024. Coverage reached more than 50 million people when Washington came online in May 2021. Delivery runs through Wireless Emergency Alerts on FEMA’s IPAWS at an estimated M 5.0 and MMI IV, through the MyShake app and through native Android operating-system integration at MMI III.

Established Japan has the longest nationwide record and the most informative failure. The Japan Meteorological Agency’s public system began in 2006–2007. Kodera and colleagues documented in 2021 what the 2011 M 9.0 Tohoku-Oki earthquake did to it: the system underpredicted ground motions for the Kanto region about 300 km southwest of the epicentre because its point-source assumption could not represent a fault hundreds of kilometres long; it missed several large aftershocks because the P-wave picking algorithm failed when onsets overlapped the coda of earlier shocks; and it overpredicted for multiple simultaneous earthquakes by mis-associating P-wave arrivals and mislocating hypocentres. The agency’s own prediction score, which had run at about 80% from October 2007 to March 2010, fell to 28% in Japanese fiscal year 2010.

Established Japan’s response was algorithmic and it is measurable. The Integrated Particle Filter method entered service on 14 December 2016 and lifted the prediction score from 69.7% to 83.6%, largely by removing over-warnings for M 2–3 events. PLUM entered service on 22 March 2018 and scored 84.6% on its own, improving timeliness by about 3.4 s on 5 of 13 test earthquakes. The S-net seafloor cable array was integrated to see subduction ruptures before they reach land. This is the clearest instance in the whole subject of a measured failure driving a measured fix.

Established The 1 January 2024 M 7.6 Noto Peninsula earthquake is the best recent test of the improved Japanese system, and it passed on accuracy while showing the blind zone. Yamada and colleagues reported in Earth, Planets and Space in 2025 that the first public warning went out 6 s after the first P-wave detections, and that the blind zone was a few tens of kilometres — small relative to the extent of strong shaking, but centred exactly on the worst of it. At the second warning, 26 of the 35 districts that should have been alerted were not; by the final warning only five false negatives remained, all at intensity 4, alongside ten false positives. Only one of the 35 districts observing intensity 4 or greater ended with an intensity error of two units or more. Residents in Nanao received the message before the P wave arrived, but only because smaller earthquakes minutes earlier had already triggered the system.

Established Mexico has the longest operational record of any public system and the starkest illustration of geometry. Suárez’s 2022 retrospective records SASMEX running continuously since August 1991, detecting roughly 9,800 earthquakes, issuing 111 alerts and producing exactly one false alert — on 16 November 1993, during the testing phase. It runs 97 accelerograph stations, broadcasts to six cities, serves about 25 million people, and reaches them through roughly 90,000 dedicated radio receivers and about 12,600 public loudspeakers in Mexico City as of 2015. For subduction earthquakes on the Pacific coast, Mexico City gets at least 50 s. For inland and intraslab earthquakes it gets 10 s or less.

Established September 2017 delivered both halves of that record within twelve days. For the 7 September M 8.2 Chiapas earthquake, sirens sounded across Mexico City roughly two minutes before shaking began. For the 19 September M 7.1 Puebla earthquake — the only event to kill anyone in Mexico City since 1985, with more than 200 dead in the city and about 150 elsewhere — the sources disagree about whether the alert preceded the shaking at all. Suárez states that the alert “was heard in Mexico City at the same time or a few seconds after the arrival of the strong seismic shaking.” The Eos account puts the alert about 5 s after P-wave arrival and roughly 20 s before the S wave. A USGS reconnaissance report gives 3–5 s after initial wave arrivals. An Issues in Science and Technology essay says 12 s before serious shaking. This brief does not average them: the honest statement is that the alert and the shaking arrived within a few seconds of each other, and which came first depended on where in the city you were standing.

Established The rest of the SASMEX record is better than the 2017 headline suggests. For the 16 February 2018 M 7.2 Oaxaca earthquake the system delivered 78 s to Mexico City, 64 s to Puebla, 44 s to Acapulco and 28 s to Oaxaca city; two days later an M 6.0 aftershock gave Mexico City about 30 s. Across four large shocks in six months, three produced 30–120 s of warning for Mexico City. The one that failed was the one that killed people, and that is not a coincidence: the events that are close enough to be lethal are the events that are close enough to be inside the blind zone.

Established China has built the largest network and its first public performance report shows the bottleneck has moved to distribution. Peng and colleagues reported in Remote Sensing in 2022 that the national system passed 15,000 stations by the end of that year — about 2,000 datum, 3,200 basic and more than 10,000 MEMS-based ordinary stations — with more than 20,000 planned and mean spacing of about 12 km in five priority zones, on an investment of 1.95 billion RMB across a five-year cycle beginning August 2015. In the 2022 M 6.1 Lushan earthquake the first alert came 5.7 s after origin with a 1.3 km location error, a sub-kilometre depth error and a 0.1 s origin-time error, but with the magnitude at M 4.2 — 1.9 units low. Five successive alerts walked the estimate up to M 6.1 by 16.5 s. The observed blind zone was about 30 km against a theoretical 10 km, and the authors attribute the gap to an alert distribution system that did not account for each terminal’s epicentral distance.

Frontier The evidence that any of this reduces harm is thin, and the thinness is specific rather than general. Cremen and Galasso’s 2020 review in Earth-Science Reviews counted EEW operating in nine countries and under test in thirteen more, and observed that few if any implemented systems trigger alarms on explicit engineering damage or loss predictions at all. Velazquez and colleagues, reviewing the technical and socio-organisational components the same year, could point to one concrete industrial instance: a Japanese electronics plant that cut its chemical supply 12 s before shaking arrived from the 2007 M 6.8 Chuetsu-oki earthquake. The BART network in the San Francisco Bay Area is wired to slow trains on a ShakeAlert message and responded correctly to the 2014 M 6.0 South Napa earthquake — which struck at 3:20 am, when no trains were running.

Frontier The loss-reduction figures in circulation are model outputs, and the models assume the thing they are asked to demonstrate. Porter’s 2016 estimate that 19,000 people in the San Francisco Bay Area could avoid injury or death assumes the whole community trained and executing drop-cover-hold-on. Strauss and Allen’s 2016 cost-benefit case is that ShakeAlert’s annual operating cost would be recouped if the system saved three lives, warned two semiconductor plants, slowed one train enough to avoid a derailment, prevented 1% of non-fatal injuries and cut gas-related fire damage by 0.25%. Every one of those five conditions is a conditional rather than an observation. The most careful treatment is Papadopoulos and colleagues in Earthquake Spectra in 2023, which estimated that an EEW system in Zurich could reduce injury risk by about 20–25% at short return periods and 15–20% at longer ones — on assumed inputs of 80% alert receipt, 65% response, 70% attempting drop-cover-hold-on with a median reaction time of 8.8 s, 25% fatality reduction and 65% injury reduction conditional on a successful drop-cover-hold-on, and a binary assumption that alerts do nothing during seven hours of sleep. The authors are explicit that these come from literature plus critical judgement, not from Swiss measurement.

Established Where human response has been measured directly, it is much worse than the models assume. Tan and colleagues’ 2022 review in Frontiers in Communication reports that in Nakayachi and colleagues’ Japanese survey (n = 359), 7% of recipients took physical protective action while 25% used the alert only to prepare themselves mentally; a Peruvian study by Fallou and colleagues (n = 2,625) likewise found most recipients taking no protective action. The review’s own summary of the field is the sentence this brief keeps coming back to: twenty papers presumed benefits, and the outcome studies found fewer people doing the prescribed thing than expected. A model that assumes 65% response against a measured 7% is not slightly optimistic; it is off by most of an order of magnitude on the one parameter the answer is most sensitive to.

Established The injury mechanism the whole case rests on is, at least, well established. More than half of injuries in the 1989 M 6.9 Loma Prieta earthquake were linked to falls, and more than half in the 1994 M 6.7 Northridge earthquake to non-structural falling hazards, with injuries alone costed at $2–3 billion. Those are the injuries drop-cover-hold-on is designed to prevent. The causal chain from alert to averted injury is therefore plausible at every link except the one nobody has instrumented: whether the alert changes what people and machines actually do, at scale, during real shaking.

Frontier The central open question is not seismological but epidemiological: does a population with warning suffer fewer injuries than an otherwise identical population without? Every operational system publishes latency, magnitude error and alert counts. None publishes an injury rate difference. The subject has a complete process metric and no outcome metric, which is the same structural defect this corpus identifies in Infrastructure Resilience and in Forecasting Replication: an elaborate optimisation literature with no error signal at the end of it.

Frontier Whether magnitude estimation can be made robust for great earthquakes in the first ten seconds is genuinely contested. PLUM’s answer is to abandon magnitude; finite-fault and real-time GNSS approaches try to measure the growing rupture directly from static displacement, which does not saturate; machine-learning approaches try to extract more from the same early P-wave window. The Chinese system’s Lushan performance — M 4.2 at 5.7 s for an M 6.1 event, correct only by 16.5 s — is a fair picture of where point-source estimation stands, and Minson and colleagues’ 2018 argument is that no amount of processing cleverness removes the underlying problem, because the information is not yet in the seismogram.

Frontier Threshold policy is an unresolved optimisation with real and asymmetric costs. The Earth’s Future analysis of 143 earthquakes shows there is no threshold that both covers everyone who needs an alert and spares everyone who does not. ShakeAlert resolves the trade by alerting low and accepting that ten of its 21 Wireless Emergency Alerts went out for earthquakes below M 5.0. Japan resolves it differently, at district granularity, and the Noto event still produced ten false positives alongside five false negatives at the final warning. Nobody has published what the optimal point is, because the loss function on the alert side has never been measured.

Frontier Offshore detection for subduction megathrusts is the most consequential open engineering question. Japan’s S-net exists; Cascadia’s equivalent largely does not, which is why the Pacific Northwest M 9 simulations put just under 60% of the population inside a timely MMI V alert rather than a higher figure. Whether a seafloor array changes that number materially, and by how much per kilometre of cable, has not been published for Cascadia.

Speculative Whether automated action or human action is the larger share of the available benefit is unresolved and the evidence tilts toward machines. Machines have no reaction-time distribution, do not sleep, and do not need to decide. Papadopoulos and colleagues assume a median human reaction of 8.8 s against warning times that are frequently under 10 s; a relay opens in milliseconds. If that asymmetry is as large as it looks, most of the realisable benefit sits in a few thousand industrial and transport installations rather than in tens of millions of phones — and that is a testable proposition, which section 7 sets out how to test.

Established The binding constraint is not a technology; it is the absence of an outcome measurement, and it is cheap to fix. Japan has run automated elevator, gas and rail responses for close to two decades; Mexico has alerted almost all public schools in Mexico City since the 1990s through about 90,000 receivers. The exposure exists. What does not exist is a registry that records which installations were wired to act, which alerts they received, and what happened to their people and equipment. Every other bottleneck in this list is downstream of that one, because without it there is no way to price a second.

Established The blind zone is a hard physical bottleneck and it is centred on the worst shaking. At 2–3 s of communication latency and 5 s of human reaction, the no-warning radius is about 25 km; purpose-built dense networks can reach 15–25 km. Areas that reach MMI VII or above — the intensities where injuries and deaths happen — typically sit inside it. The USGS states the consequence plainly with respect to the Hokkaido (41 deaths), Kumamoto (45 deaths) and Ridgecrest events: most residents in the area of strongest shaking received no warning. No engineering programme removes this.

Established Magnitude saturation is the second physical bottleneck and it bites hardest on the events that matter most. Tohoku is the canonical case: a point-source algorithm underpredicted Kanto shaking 300 km away because the fault was longer than the model allowed, and the agency’s prediction score fell to 28% for the year. Lushan is the ordinary case: 1.9 magnitude units low at first alert. PLUM and finite-fault methods are partial answers that buy accuracy at the cost of lead time.

Established The last mile is a larger share of the latency budget than most accounts admit. For Alum Rock, the gap between the fastest fifth of handsets (under 2.2 s) and the slowest fifth (over 6.0 s) is comparable to the entire detection time. At Lushan the distribution layer tripled the blind zone. Improving telemetry, alert fan-out and handset wake-up is unglamorous, invisible in the seismological literature, and probably where the next two seconds are.

Frontier Human reaction time is a bottleneck that no amount of system speed relieves. The best available model assumes a median 8.8 s to complete drop-cover-hold-on and 25 s to evacuate, against warning times that are frequently zero to 23 s at damaging intensities. Measured protective-action rates are 7% in one Japanese survey. If those numbers are right, the human channel delivers a small fraction of the modelled benefit and the machine channel delivers the rest.

Established Offshore instrumentation is a capital bottleneck with a known solution. Subduction megathrusts nucleate under water. Japan cabled its margin; the Cascadia margin is not comparably instrumented, and the M 9 population simulations reflect that gap. Seafloor cable is expensive, slow to permit and slow to lay, and it is the one component of this system whose lead time is measured in years rather than seconds.

Speculative The absence of a liability rule for automated action is a bottleneck operators will not say aloud. An elevator that stops between floors on a false alert, a surgical pause that costs a patient minutes, a train braked hard on an over-estimate: each has an identifiable defendant and no established standard of care. Japan and Mexico have decades of practice and therefore de facto norms; jurisdictions starting now have neither, and this is the sort of constraint that quietly sets the adoption rate.

Established The first dependency is a shared, auditable alert-and-action ledger, and it is a records problem rather than a science problem. Every operational system already logs every message it emits with a timestamp and a predicted intensity field. Every automated installation already logs its own trips. Joining those two logs, with the location and identity of the installation and an incident record from the operator, is the entire technical content of the decisive experiment in section 7. Nothing has to be invented; something has to be agreed.

Frontier The second is a measured, not assumed, human-response distribution. The dominant risk model in this field uses a median drop-cover-hold-on time of 8.8 s with a logarithmic standard deviation of 0.4, and an assumed 65% response probability, against a measured 7% physical-action rate in one survey. Closing that gap requires instrumented behavioural measurement during real events — accelerometers, video in consenting institutional settings, and structured post-event surveys at scale — not more modelling.

Established The third is a catalogue standard that lets systems be compared. ShakeAlert’s performance is defined against the USGS Comprehensive Catalog with a 100 km and 30 s matching rule; Japan scores against JMA intensity districts; China against its own catalogue. The published numbers — 22.6% miss, 84.6% prediction score, 111 alerts and one false alert — are not on the same scale, and no cross-system comparison in this brief should be read as if they were. Shared definitions belong in the same class of work as Metrology Infrastructure.

Frontier The decisive experiment is a matched-facility automated-action study, and it is the one result that would most change this brief’s assessment. Take facilities of the same class — hospital operating suites, semiconductor fabrication lines, elevator estates, gas distribution districts, rail sections — in the same alerting region, some wired to take automatic action on an earthquake early warning message and some not, and measure injury rates and equipment-damage rates across a decade of real earthquakes. The comparison is between facilities exposed to the same shaking, so the ground motion is a covariate rather than a confounder, and the outcome is counted in injuries and repair invoices rather than in seconds.

Established The exposure for this study already exists and has existed for years. Japan has run automated elevator, gas-shutoff and rail responses at national scale since the late 2000s; Mexico alerts almost all public schools in Mexico City; the United States has more than 50 million people under coverage and a growing set of institutional subscribers. Facilities adopt automated action for idiosyncratic reasons — a safety officer’s initiative, an insurer’s discount, a retrofit cycle — which is close to the assignment mechanism a natural experiment needs.

Frontier What is missing is the ledger, not the events. The study requires three joins that nobody has made: the alert log, which every operator keeps; the trip log of each automated installation, which the installation keeps; and the incident record, which the facility keeps and generally does not share. A mandate on operators of automated seismic actuation to register the installation and report post-event outcomes would create the dataset within one seismic cycle of a moderately active region.

Speculative The expected effect size is large enough to be detectable and small enough to be missed if the study is underpowered. If the modelled 20–25% injury-risk reduction for a whole population is anywhere near right, the effect in a facility that acts automatically should be larger, because the machine channel does not depend on the 7% of people who take physical protective action. If the measured effect in automated facilities is indistinguishable from zero, the case for public alerting as an injury-reduction measure collapses and the case for it as an industrial-process measure collapses with it.

Frontier The second experiment is a randomised message-format trial inside an existing alerting population. The systems already deliver tens of millions of messages; the content, wording, sound and intensity threshold of those messages have never been randomised at scale against a behavioural outcome. Given a measured protective-action rate of 7% in one survey, even a doubling would be worth more than several seconds of engineering latency, and it costs nothing but institutional nerve.

Established The third is already running and will report on nature’s schedule. A Cascadia megathrust rupture is the test of every claim in section 3 about great earthquakes, and the population-based simulation that stands in for it — 30 synthetic M 9 scenarios, just under 60% of the Pacific Northwest population alerted before MMI V shaking — is a prediction with a date nobody controls. The same is true of the next great subduction earthquake off Mexico or Japan.

Frontier The fourth is a delivery-latency audit, and it would settle where the next second should be bought. Instrument the path from detection to actuation end to end, at the handset and at the relay, across a full alert population. The Alum Rock figures — 20% of phones under 2.2 s, 80% under 6.0 s — and the Lushan finding that distribution tripled the blind zone both suggest the answer is in software, not seismometers, and that hypothesis is cheap to falsify.

Speculative The fifth is a cost-per-second-of-warning curve for a real jurisdiction. Li, Sornette and Feng produced the only published example of the genre: about 3,000 km² of small-blind-zone coverage per 1,000,000 RMB of densification beyond the Chinese plan. The equivalent for Cascadia, comparing land stations against seafloor cable against telemetry upgrades against distribution software, does not exist and would take one analyst-year given data operators already hold.

Frontier The sixth is an alert-harm study, which nobody wants to run and which the ethics of the field require. Alerts can cause injury: people run, fall, leave buildings mid-shaking, or brake hard on a motorway. No operational system publishes an alert-attributable injury count. The same matched-facility ledger that measures benefit would measure this, and a system that cannot say whether its harms are smaller than its benefits is asserting rather than reporting.

Established What none of these needs is new physics. Every experiment above runs on instrumentation that exists, on populations already alerted, and on events that will happen whether or not anyone is measuring. The obstacle is that the institutions holding the three logs have no obligation and no incentive to join them, which section 10 treats as the governing constraint on the whole subject.

Established The sensor layer is the mature part and its specification is now conventional. ShakeAlert runs 1,509–1,553 broadband and strong-motion stations against a 1,675 target; China runs more than 15,000 of three classes, including more than 10,000 MEMS units, toward more than 20,000; SASMEX runs 97 accelerographs. The design variable is spacing: about 12 km in China’s five priority zones, and the blind-zone radius follows from it almost mechanically. MEMS accelerometers are what made the Chinese station count possible, and they are the reason station count is no longer the expensive part.

Established The public-delivery layer is three parallel channels with different thresholds and different failure modes. Wireless Emergency Alerts through FEMA’s IPAWS fire at an estimated M 5.0 and MMI IV and reach everyone with a phone but cannot be personalised; the MyShake application and native Android operating-system integration fire at MMI III and can be; Mexico uses about 12,600 municipal loudspeakers and about 90,000 dedicated receivers, which are slower to deploy but do not depend on a cellular network that an earthquake may have just damaged. Redundancy across physical layers is the design principle, and the loudspeaker is not an anachronism.

Frontier The actuation layer is where the engineering is least standardised and most of the value probably sits. Elevators stopping at the nearest floor and opening doors, gas lines closing, trains braking, manufacturing processes safing, data committed to disk, traffic signals turning red before bridges and tunnels: each needs a controller that accepts a message, applies a local intensity threshold, and fails safe when the message does not arrive. There is no common protocol profile across jurisdictions, no published conformance test, and no standard for what a controller does on a message that later gets downgraded — and messages do get downgraded, by 2.3 magnitude units in the Truckee case.

Frontier The single-facility engineering question is whether to act on the regional message or on a local sensor. Minson and colleagues analysed this for a Northern California rail system in 2021 and reached a result that should be better known: on-site warning triggering at a 120 gal threshold outperformed source-parameter warning regardless of false-alarm tolerance, and the dominant benefit came from stopping trains before they reached track that had already been damaged, rather than from slowing them before shaking arrived. That inverts the usual story. Where it generalises, the economic case for EEW is a post-event routing case as much as a pre-event braking one.

Established Everything here is a dependency of, not a substitute for, structural engineering. A few seconds cannot change how a building behaves; it can only change where people and machines are when it behaves that way. The same reasoning that governs Infrastructure Resilience applies: warning is a complement to hardening with a different cost curve and a different failure mode, and treating it as a substitute is the error to avoid.

Established Tsunami warning is the closest neighbour and shares the expensive hardware. Both need offshore instrumentation on subduction margins, both need rapid magnitude estimation for great earthquakes, and both were reshaped by the 2011 Tohoku event. The cadences differ by three orders of magnitude — seconds against tens of minutes — which is why tsunami warning can afford accurate magnitudes and earthquake early warning cannot.

Frontier Distributed acoustic sensing on existing fibre is the most plausible step-change in station density. Turning installed telecommunications fibre into a dense array of strain sensors would change the economics of spacing in exactly the direction the blind-zone arithmetic rewards. Whether the noise characteristics support reliable sub-second P-wave picking in urban conduits is not settled in the sources this brief reached, and no operational system in it reports using DAS in its alert loop.

Frontier Machine-learning nowcasting has a structural parallel in weather. The problem shape — take a partial early observation, predict a field a few seconds or hours ahead, and be scored on it — is the one treated in AI Weather Prediction, and the contrast is instructive: weather forecasting has a scoring culture with decades of verification statistics, and earthquake early warning has process metrics and no outcome verification at all.

Established The operating model is a public agency running a utility-grade service on annual appropriations, and that is a poor match to the asset. The United States system took $161.8 million between FY2006 and FY2024, with annual appropriations rising from $6.5 million to $28.6 million, to reach 1,553 of 1,675 stations by December 2024. A seismic network is a perpetual operating commitment — power, telemetry, calibration, replacement — funded by a mechanism designed for projects that finish.

Established A performance standard exists and the system is outside it. ShakeAlert carries a standard of a 10% or lower miss rate; the Congressional Research Service reports an outturn of 22.6%. That is the right kind of institutional artefact — a published target and a published outturn — and it is rare enough in hazard programmes to be worth naming as good practice even while the number is bad. It is also fragile: the misses were concentrated at network edges, a fact that lives in the source paper and not in the standard.

Frontier No jurisdiction has a settled liability regime for automated seismic action. Who is responsible when an elevator stops between floors on an alert for an earthquake that turns out to be M 3.7, or when a surgical pause costs a patient two minutes? Japan and Mexico have decades of practice and therefore working norms; everyone else is improvising. Until there is a standard of care, the rational choice for a risk-averse facility manager is to subscribe to alerts and not wire them to anything, which is precisely the choice that keeps the decisive experiment unrunnable.

Frontier The missing institution is a registry and reporting duty for automated actuation. It is a small body: a register of installations, a standard post-event return, and an obligation on the alert operator to publish the matched alert log. It would cost a fraction of one year’s station maintenance and it is the only route to the measurement in section 7. That such a body does not exist anywhere after thirty-five years of operational warning is the central institutional finding of this brief.

Frontier The public-communication function is chronically underfunded relative to the sensing function. Almost every dollar in this subject buys instrumentation, and the measured bottleneck is behavioural: 7% physical protective action in one survey against 65% assumed response in the leading risk model. The same institutional asymmetry appears across hazard science, which is why Crisis Science Institutions is the right neighbour for this problem rather than a seismological one.

Established The system’s central equity problem is geometric and cannot be legislated away: the people nearest the rupture get nothing. The blind zone is where MMI VII and above occur, and it is where the injuries and deaths occur. The USGS names Hokkaido (41 deaths), Kumamoto (45 deaths) and Ridgecrest as cases where most residents in the strongest shaking received no warning. Every benefit statistic for this technology is therefore computed over a population that excludes the worst-affected, and any published average that does not say so is misleading by construction.

Established September 2017 in Mexico City is the ethical case study, and it cuts both ways. The one event that killed people in the city since 1985 was the one where the alert did not arrive usefully ahead of the shaking. It was also preceded that morning by a national commemorative drill at 11:00, and the USGS reconnaissance team found both that the coincidence caused confusion and doubt when the real alert sounded two hours later, and that people reacted faster because the drill had reminded them what to do. The same fact supports opposite conclusions depending on which respondents you ask, and no one has measured which effect dominated.

Frontier Alerts can injure people, and no operator reports how often. Running, falling, leaving a building mid-shaking and braking hard in traffic are all documented responses to sudden alerts in the broader warning literature, and each has an injury mechanism. The risk model that gives EEW a 20–25% injury reduction for Zurich assumes 30% of responders attempt evacuation with a median 25 s completion time — against warning times often under 20 s. Evacuation that does not finish before the shaking starts is the highest-risk behaviour available, and the model treats it as an input rather than as a harm to be counted.

Established Access is unequal in ways that track the usual lines. Phone-based delivery requires a modern handset, a charged battery, a working cell site and a language the message is offered in. Mexico’s loudspeaker and dedicated-receiver model reaches people without any of those, and its coverage of almost all public schools in Mexico City is a distributional choice made explicitly. Which delivery mix a jurisdiction picks is a decision about who is protected, and it is usually made as a procurement decision.

Speculative Consent is absent by design and that is defensible, but it should be stated. Public alerting to every handset in a region is a compulsory intervention in millions of lives, several times a year in an active region, on the basis of a modelled benefit that has never been measured. The honest position is that the intervention is cheap, the plausible harm is small, and the evidence obligation therefore falls on whoever wants to expand it rather than on whoever wants to stop it — which makes the measurement in section 7 an ethical requirement rather than an academic one.

Frontier Earthquake early warning is the first deployed instance of a general capability: a civilisation-scale reflex operating below the timescale of human decision. The system detects, decides and actuates in under ten seconds across territories of millions of square kilometres, and the human being is the slowest component in it by an order of magnitude — 8.8 s of assumed median reaction against milliseconds for a relay. Whatever else it is, it is a working demonstration that a public institution can hold a continent-scale automated loop closed for decades.

Established The precedent has already travelled and its scale is measurable. Thirty-five years of Mexican operation, roughly two decades of Japanese nationwide operation, seven years of United States public alerting covering more than 50 million people, and a Chinese network past 15,000 stations built for about 1.95 billion RMB. Volcanic, landslide and flash-flood warning all inherit the architecture. The reusable asset is not the seismometer; it is the social licence to interrupt everybody at once on a machine’s judgement.

Speculative The most consequential long-run effect may be on what gets automated, not on what gets warned. If the matched-facility study finds the benefit concentrated in automated actuation rather than in human response, the rational policy shifts from alerting populations to wiring infrastructure — elevator estates, gas districts, rail sections, fabrication lines. That is a smaller, duller and much cheaper programme than public alerting, and it is politically far less attractive, which is one reason nobody has been eager to run the study.

Frontier The hard limit is that this technology moves people and machines, not buildings. The dominant determinant of earthquake deaths remains building stock, and no quantity of seconds changes a masonry wall. A civilisation that treats warning as a substitute for seismic codes and retrofits has bought a cheap thing instead of an expensive one that works, which is the substitution error also identified in Civilization Resilience Planning.

These horizons track measurement and institutional adoption rather than detection physics, which is close to its floor already. The near horizons are stated with more confidence because nothing in them requires an invention.

• 10 yr: Frontier The United States network completes at or near 1,675 stations and the miss rate against the 10% standard is either met or formally revised; Cascadia gains some offshore capability and the just-under-60% figure for MMI V alerting in M 9 scenarios is recomputed against it. Speculative The first matched-facility outcome study reports, most likely from Japan, using elevator and gas-shutoff estates that have been operating since the 2000s. Frontier Distribution latency, not sensor density, becomes the acknowledged frontier after an operator publishes an end-to-end audit showing where the seconds actually go. Speculative At least one jurisdiction adopts a registration duty for automated seismic actuation, and it happens after an incident rather than before one.
• 25 yr: Speculative A published injury-rate difference between alerted and unalerted facilities exists for at least one country, and it is smaller than the 20–25% modelled figure because the human response channel underperforms its assumptions. Speculative Automated actuation rather than public alerting becomes the primary justification for funding, shifting the programme from communications to industrial control. Frontier Real-time geodesy and wavefield methods jointly remove magnitude saturation as a practical failure mode for great earthquakes, which will be demonstrated by an actual megathrust rather than by simulation. Handwave Expect the first such demonstration to be contested on the grounds that the comparison case was a different rupture.
• 50 yr: Speculative Seconds-scale automated hazard response is an ordinary building service in seismic regions, specified in codes alongside sprinklers, and the marginal argument is about conformance testing rather than about whether to do it. Speculative The blind zone is unchanged at 15–25 km because it is set by wave physics, and the public understanding of the system has finally caught up to that fact. Handwave Whether public alerting survives as a mass-notification product, or narrows to institutional actuation plus a shrinking consumer channel, depends entirely on an outcome number that does not yet exist.
• 100 / 250+ yr: Handwave The interesting question is not seismological. If a civilisation retains continuous continental instrumentation and automated actuation for centuries, the asset is the institutional habit of holding a fast closed loop across jurisdictions, and its value shows up in hazards nobody currently associates with seismology. Handwave The counter-case is equally available: warning networks are among the first things a strained state stops maintaining, because their benefit is invisible in every year that nothing happens, a failure mode this map treats at length in Infrastructure Resilience.

• Depends on Less than a reader might expect, and that is the finding. No result in another brief on this map blocks the physics: detection, telemetry and actuation are all deployed and measured. What it does wait on is the measurement culture described in Infrastructure Resilience — that brief’s central complaint, an elaborate optimisation literature with no published outturn distribution, is this subject’s complaint word for word — and on the scoring discipline that Forecasting Replication treats generally. It also waits on the comparability work in Metrology Infrastructure, because the published figures from the United States, Japan, Mexico and China are not currently on the same scale.
• Requires (not on this map) A register of installations wired to act automatically on an alert, with a post-event reporting duty, which is the whole technical content of the decisive experiment. A settled standard of care for automated seismic action, without which risk-averse facilities subscribe to alerts and wire them to nothing. A measured injury-rate difference between facilities that act automatically and matched facilities that do not, which no country has produced. A measured protective-action rate during real shaking, to replace the 65% assumed in the leading risk model with something closer to the 7% one survey found. Station spacing below about 20 km through populated fault zones, which is an industrial deployment problem rather than a research one. Cabled seafloor instrumentation on the Cascadia margin, on the Japanese S-net model, which is the only route to improving the just-under-60% figure for M 9 alerting. And a market in which a facility can buy a specified number of seconds under contract, which would create the first honest price for the product.
• Enables A seconds-scale automated reflex layer that other hazards can reuse — volcanic, landslide, flash-flood and industrial-process warning all inherit the architecture once the actuation interface is standardised. It also enables something narrower: an evidential base for hazard-warning policy generally, because this is the only warning technology whose latency budget is short enough for the counterfactual to be clean. If the matched-facility study reports, it becomes the template for every other warning system that currently reports process metrics and no outcomes.
• Adjacent Tsunami warning, which shares the offshore hardware and the great-earthquake magnitude problem; structural health monitoring, which shares the sensors and closes the loop after the event; the automated-actuation estate examined in Robotics in Infrastructure; the nowcasting and verification culture of AI Weather Prediction; the multi-hazard alerting question in Compound Climate Hazards; and the institutional design problem set out in Crisis Science Institutions.

Established “Earthquake early warning predicts earthquakes.” It does not, and the distinction is the whole subject. Every operational system waits for a rupture to start and then races its own message against the slower seismic waves. Nothing in ShakeAlert, JMA, SASMEX or the Chinese network says anything about an earthquake that has not begun. The claim circulates because the name invites it, and because prediction is what people want; the correct response is that the systems have a measured detection latency of 4–8 s from origin time, which is only meaningful for an event that has already happened.

Established “You get about a minute of warning.” That figure comes from one city and one geometry. Mexico City receives at least 50 s for subduction earthquakes because it sits several hundred kilometres from the source region; it received 78 s in February 2018 and roughly two minutes in September 2017. For inland and intraslab events the same system delivers 10 s or less. In California, the delivered warning at MMI VI sites was 11–23 s for Petrolia in 2021 and 0 to 23 s for Ferndale in 2022. A minute is not the typical case; it is the best case of the most favourably situated megacity on earth.

Established “The Mexican system failed in 2017.” It performed to specification and the specification was beaten by geometry. On 7 September 2017 the system gave Mexico City about two minutes for an M 8.2. On 19 September the rupture was roughly a hundred kilometres away rather than four hundred, and the alert and the shaking arrived within a few seconds of one another — sources place the alert anywhere from 12 s before serious shaking to a few seconds after it. The lesson is not that the system was broken but that a nearby earthquake is inside the blind zone by definition.

Frontier “More sensors will eventually eliminate the blind zone.” They will shrink it and cannot close it. The floor is set by the time for waves to reach the nearest instrument plus 2–3 s of communication and about 5 s of human reaction, which puts the radius near 25 km today and 15–25 km for purpose-built dense networks. The marginal return is real and has been costed — about 3,000 km² of additional small-blind-zone coverage per 1,000,000 RMB in the Chinese case — but it is a return on area, not on the epicentral region itself.

Frontier “Earthquake early warning saves lives, this is established.” It is not established; it is modelled. The circulating figures — 19,000 injuries and deaths avoided in the San Francisco Bay Area, injuries reduced by more than half, a 20–25% injury-risk reduction for Zurich — are model outputs conditioned on assumed behaviour. The best of them assumes 80% alert receipt, 65% response and a 65% injury reduction conditional on a successful drop-cover-hold-on, all sourced from literature and judgement rather than from local measurement. Against that, the measured physical-protective-action rate in one Japanese survey of 359 people was 7%. The claim may well be true. It has not been demonstrated, and section 7 says what would demonstrate it.

Established “False alarms are the main problem.” On the published record they are rare, and the real problem is the opposite. SASMEX has issued 111 alerts in thirty-plus years with one false alert, in 1993, during testing. ShakeAlert produced seven events categorised false in four years, of which six were real earthquakes badly located at network edges. The genuine error mode is the miss and the underestimate: five earthquakes had magnitudes underestimated below the alert threshold, and the miss rate of 22.6% sits above a 10% standard. Systems fail quiet, not loud.

Established “Japan’s system handled Tohoku.” It handled the detection and failed the size. The 2011 M 9.0 rupture underpredicted ground motions in Kanto 300 km away because the point-source assumption could not represent a fault of that length; the aftershock sequence defeated the P-wave picker when onsets overlapped earlier coda; and simultaneous events produced over-warnings through mis-association. The agency’s prediction score fell from about 80% to 28% for the fiscal year. The fixes — IPF in December 2016, PLUM in March 2018, S-net integration — are the field’s best documented case of a failure being measured and then engineered away, and quoting the recovery without the collapse misrepresents both.

Handwave “Machine learning will soon give minutes of warning.” This one is asserted rather than argued, and the physics says otherwise. Minson and colleagues’ 2018 result is that the time before strong ground motion can be expected at a site is bounded below by the rupture’s own evolution, because magnitude is not predictable from the first arrivals of a rupture that has not finished growing. A better algorithm can extract more from the available seismogram; it cannot read a part of the rupture that has not happened yet. Claims of minutes-scale warning for a nearby earthquake are claims about prediction wearing a detection costume.