<!-- ARTICLE_BODY_START -->
A warning about an unfinished earthquake
A 2025 evaluation of ShakeAlert's third major software generation described improvements in a particularly difficult task: estimating the consequences of a large earthquake before the earthquake has finished happening. The changes included geodetic measurements of ground displacement, revised source algorithms and treatment of local ground response. The researchers reported useful warning times at some locations during actual moderate earthquakes and examined larger events through offline tests. Those are different kinds of evidence, not interchangeable demonstrations of future performance. McGuire and colleagues' 2025 evaluation.
The distinction matters because earthquake early warning is often pictured as a race between a fast electronic message and a slow seismic wave. That picture captures an opportunity but omits the hardest competitor: the time required for the earthquake to reveal what it is becoming. A processor can calculate quickly and still lack the observations needed to distinguish a short rupture from the beginning of a much larger one.
There is no single quantity called the warning time of an earthquake. There is a warning time for a particular location, a particular shaking threshold and a particular information product. A message produced before a strong pulse arrives can be valuable even if a weaker wave has already passed. Conversely, an excellent final magnitude estimate can be useless for a location whose strongest motion occurred several seconds earlier.
This article follows the scientific estimate from the first arriving waves to its eventual evaluation. It is not a guide to emergency behavior or a promise that an alert will precede shaking. Its numerical examples are deliberately simplified, hypothetical calculations, not forecasts for named towns. Their purpose is to expose which seconds come from geometry, which are consumed by observation and which cannot be recovered merely by buying a faster computer.
Start with two paths, not one distance
The physical opportunity begins with differences in seismic-wave speed. Compressional P waves generally arrive before the slower shear waves and surface waves that commonly carry stronger shaking. An observing network can detect an earthquake already underway and communicate information ahead of some of the later motion. That is early warning, not prediction of an earthquake before it starts. Actual wave speeds depend on the materials and paths involved. USGS explanation of the physical basis.
A useful first calculation needs two distances. One runs from the source to the station or stations that supply sufficient evidence. The other runs from the source to the place being warned. Substituting the user's distance for the detector's distance accidentally describes a different system.
Consider an invented, uniform Earth in which P waves travel at six kilometers per second and S waves at three and a half. Ignore depth, curved paths, rupture length and surface waves for now. Put the effective detecting station twenty kilometers from the source. Its first P wave arrives about 3.33 seconds after origin. Suppose collecting and processing enough information then takes another 2.5 seconds. The hypothetical source estimate is ready at 5.83 seconds.
At a site sixty kilometers from the source, the modeled S arrival is about 17.14 seconds after origin. Subtracting the source-estimate time leaves approximately 11.31 seconds. That is a source-product lead time under these assumptions, not a measured phone warning and not a guaranteed interval before damage. Any subsequent delivery or action delay would consume part of it.
Move the target site to fifteen kilometers while leaving the detecting geometry unchanged. Its modeled S arrival becomes 4.29 seconds. The estimate appears roughly 1.55 seconds too late. Nothing contradictory happened: the same detection supported a positive lead time at one site and a negative lead time at another.
Now move the effective detector from twenty to forty kilometers away. With everything else fixed, its P arrival is delayed by another 3.33 seconds. The sixty-kilometer target loses that same amount of lead time. The change is not a failure of telecommunications; it is additional travel through the Earth before the required observation exists.
These calculations also reveal why an early-warning blind zone is not a permanent circle with one universal radius. Its extent depends on detection requirements, station positions, source depth, wave paths and the threshold used to define the arrival that matters. In this toy model, the zero-lead distance is the assumed S speed multiplied by the 5.83-second estimate time, or about 20.42 kilometers. Change the assumptions and the boundary changes. A mapped operational blind zone requires a substantially richer calculation.
The fourth station can matter more than the nearest
A station recording motion does not automatically establish a reliable earthquake origin. A network must distinguish a propagating seismic signal from local disturbance and determine which observations belong together. Multiple arrival times constrain a source differently from one unusually clear waveform.
ShakeAlert's published algorithm description says EPIC uses P-wave observations and requires triggers from at least four stations, together with a nearby-station consistency condition. Its estimates are revised as observations accumulate. The point is not that four is a universal scientific constant. It is that a network detection rule imposes an evidence requirement, and the earliest individual recording may not satisfy it. ShakeAlert's algorithm documentation.
Imagine four hypothetical stations whose usable P-wave observations become available at two, three, four and nine seconds after origin. A rule requiring all four cannot act at the average arrival time of 4.5 seconds. It must wait for the last required observation. Adding a fifth station that supplies independent evidence at five seconds could change the relevant waiting time, even if that new station is farther from the source than the nearest existing instrument.
This is why station-count headlines can conceal important differences. Ten stations placed along nearly the same direction do not necessarily constrain a source as effectively as a smaller number surrounding it. If the network sees an offshore event from land, candidate locations farther seaward may resemble one another in their predicted arrival pattern. More data help only to the extent that they distinguish among the competing possibilities.
Location and origin time are coupled. An assumed source farther from the instruments can sometimes be partly compensated by an earlier origin. A source solution is therefore not simply the intersection of perfectly measured circles. It involves uncertain arrival picks, an assumed velocity structure and incomplete spatial coverage.
For evaluation, the informative object is the sequence of estimated origins, not just the final dot. Did the first location place the source on the wrong side of the network? When did new stations resolve that ambiguity? How much of the magnitude change came from correcting distance rather than observing additional rupture? Those questions connect a system's output to an identifiable physical or observational cause.
The first seconds do not contain the final answer
A large earthquake is an extended failure process, not a completed object instantly broadcasting a label. The earliest available waveform describes a limited part of that process. Its usefulness does not imply that it uniquely determines the final magnitude.
A 2019 study analyzed more than 130,000 vertical-component waveforms from earthquakes near Japan. It found that the relationship between early P-wave displacement and magnitude saturated in a way that depended on the measurement window. The authors linked this behavior to a nondeterministic model of rupture growth and developed time-dependent uncertainty estimates. The finding concerns the information in a short observation window; it is not a claim that every large earthquake is undetectable. Trugman and colleagues on time-dependent saturation.
A hypothetical comparison makes the consequence concrete. Imagine two ruptures with observationally similar first two seconds. One stops soon afterward. The other continues across a larger fault area. An algorithm restricted to those same first two seconds cannot reliably distinguish them by consulting later measurements that do not yet exist. It can assign probabilities based on previous experience, but a probability about continuation is not an observation of completed rupture.
Waiting changes the problem. At six seconds, the continuing event may have produced additional motion that the smaller one did not. The estimate can become more informative, but some locations have spent four additional seconds approaching strong shaking. The scientific challenge is not to eliminate this trade-off through rhetoric. It is to quantify how much uncertainty remains at each elapsed time.
Magnitude updates should be interpreted accordingly. An increase can reflect continuing rupture, improved location, additional station coverage or a revised assessment of earlier measurements. Those causes have different implications for diagnosing an algorithm. Calling every upward revision an initial mistake overlooks the distinction between an inaccurate estimate of an already observable quantity and an estimate made before the relevant evidence was available.
There is also a difference between a detector's minimum usable signal and an estimator's upper useful range. An instrument can record an unmistakable earthquake while the particular short-window relationship used to estimate magnitude has become insensitive to further growth. Solving the detection problem does not automatically solve the magnitude problem. A credible performance study examines both, at the times each result could actually have been produced.
A rupture becomes a line, then something more complicated
The distance to an epicenter is not always the distance that best describes exposure to a large earthquake. As rupture extends, another part of the fault may become much closer to a target site than the point where failure began. A source model that remains a single point can therefore misrepresent the geography of shaking even when its origin is well located.
FinDer addresses this class of problem by matching the observed spatial pattern of acceleration to templates representing extended sources. The operational description distinguishes its finite-rupture information from EPIC's point-source estimate. A line is still a simplification, but it carries information about extent and orientation that a point lacks. The complementary source models in ShakeAlert.
Take another explicitly hypothetical geometry. A target sits ten kilometers perpendicular to the far end of a straight, sixty-kilometer fault segment. If rupture began at the opposite end, the target is roughly sixty-one kilometers from that origin but only ten kilometers from the nearest part of the eventual rupture. Neither distance alone supplies the arrival time: the near segment must first rupture, and then its waves must travel to the site.
Suppose, solely for illustration, rupture advances along the segment at three kilometers per second. Reaching the far end takes twenty seconds. A wave traveling the remaining ten kilometers at three and a half kilometers per second adds approximately 2.86 seconds. That path produces an arrival around 22.86 seconds after initiation. The example does not predict an earthquake's strongest pulse. It demonstrates why rupture timing and wave propagation must be kept separate.
Research on richer source models also shows why more parameters do not automatically mean more reliable estimates. In a 2021 playback study of FinDer-related methods, a more elaborate rupture-length prediction substantially overshot the final Ridgecrest length at one stage, while another estimate was closer. The authors discussed trade-offs among inferred slip, magnitude and length. This was a research comparison, not evidence that every proposed extension was deployed operationally. FinDerS and FinDerS+ study.
The practical scientific question is which additional parameter becomes identifiable early enough to improve the forecast. A highly detailed rupture description obtained after the strongest motion is excellent for later analysis but cannot be retrospectively credited as an early warning.
Why satellite positioning enters a seismic problem
Seismic instruments and high-rate satellite-positioning measurements observe related motion through different measurement systems. Their differences can be useful when a source grows beyond the range best characterized by a short seismic window. They do not make one another unnecessary.
A 2023 study adapted peak-ground-displacement scaling from real-time Global Navigation Satellite System observations for ShakeAlert. Retrospective tests supported its contribution to magnitude estimation for large events, while the authors also identified drifting position noise, station distribution and minimum usable event size as constraints. They introduced controls on when estimates were admitted and updated. This is a bounded result about a particular method, not a claim that any satellite receiver provides an earthquake magnitude. Murray and colleagues' geodetic magnitude study.
One way to understand the complementarity is to separate dynamic motion from displacement of the ground's reference position. A sensitive seismic channel may resolve small, rapid motions exceptionally well. A suitable geodetic solution can contribute information about larger displacements without relying on exactly the same short-window magnitude relationship. The usefulness of either observation still depends on noise, timing and spatial position.
An invented quality-control example illustrates why admission rules matter. Suppose a displacement series wanders slowly by an amount comparable to the signal expected from a moderate event. Taking the largest excursion in an ever-longer window will tend to preserve an unusually large value once it occurs. That maximum is not self-authenticating evidence of a larger rupture. The processing system needs a reason to attribute the displacement to the detected earthquake rather than to the behavior of the measurement series.
Combining estimates also requires attention to shared assumptions. Two algorithms may use different instruments but inherit the same source location. If that location is wrong, agreement between their magnitude estimates does not necessarily provide two independent confirmations. Conversely, disagreement may reveal a meaningful difference in sensitivity rather than a simple malfunction.
The most revealing diagnostic is a time-aligned record of when each method became informative, what uncertainty it carried and how its inclusion changed the forecast. A final combined magnitude alone cannot show whether an additional data stream improved the seconds that mattered.
Source accuracy is not shaking accuracy
Once an earthquake's source is estimated, a second problem begins: forecasting motion at a particular site. Magnitude does not specify a unique acceleration history everywhere. Source geometry, propagation and local conditions intervene between the earthquake and the measurement of concern.
A study of three-dimensional finite-fault information found that the value of richer source geometry depended on tectonic setting, faulting style and alert threshold. Its scenario calculations generally improved alert-region accuracy when the three-dimensional source representation had the correct magnitude and mechanism. That qualification is essential: a more detailed representation is not automatically a more accurate one. Murray and colleagues' finite-source investigation.
Consider two invented sites at the same distance from a source. A distance-only forecast gives them the same central estimate. Their observed motions need not match, and a difference does not by itself prove that the source magnitude was wrong. Evaluating the source algorithm solely through one site's shaking can therefore mix errors from different stages of the prediction.
The forecast target must also be specified. Peak ground acceleration, peak ground velocity and a response measure relevant to a particular structure are not synonyms. Nor is a ground-level quantity identical to the motion of a high floor. A 2022 study using recordings from instrumented buildings examined how floor acceleration estimation depended on structural response and the accuracy of the input ground-motion prediction. It did not establish a universal conversion from a regional warning map to every room in a tall building. Ghahari and colleagues' building-response study.
For an original evaluation design, these distinctions suggest three separate comparisons. First compare estimated source parameters with a later reference solution. Then compare the predicted ground-motion quantity with suitable instrument observations. Finally, where relevant, compare a structural-response estimate with measurements from that structure. A result at one level should not silently stand in for the others.
The separation is not pedantry. It tells researchers where an improvement belongs. Better earthquake location will not necessarily correct a local response model; better building characterization will not recover a signal that reached the processing center too late. A complete error budget identifies the physical stage responsible rather than attaching every mismatch to the warning system as a single opaque object.
A timestamp can be precise and still be wrong
The interval between samples, the timestamp attached to a sample and the time it reaches a processing center are separate properties. A finely sampled waveform does not guarantee an accurate clock, and an accurate clock does not guarantee prompt transmission. Those distinctions become important when a scientific claim rests on differences of a few seconds.
The international miniSEED documentation provides a concrete example of how waveform formats can represent timing information. Its reserved headers include a correction applied to record start time and a maximum estimated timing error in seconds. The existence of those fields does not prove that every stream supplies them or that every early-warning implementation consumes them correctly. It shows that uncertainty in recorded time is a recognized property of the data, not merely a concern invented after an unfavorable result. FDSN's timing-header definitions.
Suppose two hypothetical stations truly observe corresponding arrivals at four and five seconds after origin. Their physical difference is one second. If the second station's clock is incorrectly advanced by half a second, its reported arrival becomes 5.5 seconds. An algorithm unaware of the error sees a difference of 1.5 seconds and may try to explain the extra interval through source position or the assumed propagation model. Transmitting that mistaken timestamp instantaneously does not make the observation correct.
By contrast, suppose both timestamps are accurate but the second packet arrives three seconds late. A later analyst can still recover the correct physical arrival difference. The real-time estimator may nevertheless have lacked the second observation when it was needed. Clock error changes what the observation says; telemetry delay changes when the estimator can know it. Combining them into one undifferentiated latency number obscures the repair required.
Wave identification adds another distinction. A perfectly timed sample can belong to the wrong inferred phase, or a gradual onset can make the pick uncertain. An evaluator should not attribute every travel-time residual to the clock. The residual is evidence of disagreement among observation, identification and model, not an automatic diagnosis of its cause.
Propagation assumptions also influence timing. The Swiss Seismological Service explains that locating an event combines arrival differences with a velocity model appropriate to the region. That description concerns its own earthquake-location procedure, not a claim that its detection rules are ShakeAlert's rules. SED on locating earthquake origins.
As another invented calculation, a sixty-kilometer path takes ten seconds at six kilometers per second but approximately 10.91 seconds at five and a half. The 0.91-second difference comes from the assumed propagation speed, not from processor performance. Real location algorithms do not reduce the Earth to this one path; the arithmetic simply demonstrates the scale of a timing discrepancy that can arise before any software delay is considered.
A careful replay therefore preserves both versions of time: when a signal is inferred to have occurred at an instrument and when the processing system actually received the associated information. If clocks were corrected retrospectively, that correction belongs in the experiment description. Otherwise a test may inadvertently improve both the data and their availability while crediting the entire gain to a new detection algorithm.
The archive can be complete while the warning was late
The 2019 Ridgecrest sequence supplied a particularly instructive test of network performance. A study of Southern California's waveform telemetry found that some data arrived too late for an early-warning algorithm even though the waveforms eventually reached the archive. Average bandwidth could be adequate while peak demand exceeded the capacity of particular links. The study documented substantial latency on some radio paths rather than treating all communications as equivalent. Stubailo and colleagues' telemetry analysis.
That distinction is specific to real-time science. For a later catalog, a waveform arriving a minute late can still be useful. For an estimate needed in the next three seconds, its eventual arrival does not repair its absence. Completeness measured at the end of a day answers a different question from availability during the earthquake.
Imagine a hypothetical instrument producing a usable sample at time four, with the processing center receiving it at time twelve. A retrospective replay that feeds the sample to an algorithm at time four has supplied eight seconds of information the live system did not possess. Such a replay may isolate algorithmic capability, but it is not an end-to-end reconstruction unless the artificial timing is made explicit.
The reverse mistake is possible too. Reproducing historical latency when comparing two algorithms may reveal which method was more resilient to that particular network failure, not which would perform better under every future network. Both comparisons can be informative if the question is stated before the result is interpreted.
Latency should therefore be treated as a distribution attached to observations, not merely an average added once at the end. A single delayed station can be irrelevant after several others have supplied equivalent evidence, or decisive if it occupies a unique direction. The timing penalty depends on which information is missing.
An original test matrix would vary source location, station availability and delay patterns separately. It would include correlated disruptions that remove several stations sharing a communication path, not only independent random dropouts. The purpose is to discover whether apparent redundancy survives the kinds of failures that actually deprive a source estimate of spatial information.
The threshold changes the clock
An early-warning system does not simply ask whether an earthquake exists. It asks whether the information available now supports a forecast that a specified level of shaking will be exceeded somewhere. Requiring a higher forecast threshold can delay the moment when the system has enough evidence to cross it.
A 2018 analysis examined an idealized zero-latency system while retaining the time needed for rupture growth and wave travel. Under its assumptions, lower alerting thresholds yielded more robust and longer warnings than higher thresholds. The result demonstrates that eliminating technical delay does not eliminate every limit on warning time. It should not be read as a performance forecast for every modern implementation. Minson and colleagues on physical timing limits.
For a hypothetical site, suppose an evolving forecast crosses a lower criterion at six seconds and a higher criterion at eleven seconds. Suppose the observed motion exceeds a separately defined consequential level at fourteen seconds. The corresponding source-product lead times are eight and three seconds. Both calculations can be correct; they describe different policies applied to the same sequence.
This is why warning times should never be quoted without their starting and ending events. Time from earthquake origin to first estimate is not time from estimate to the arrival of strong motion. First P arrival is not necessarily the first crossing of the threshold relevant to an evaluation. Maximum shaking is another endpoint again.
Threshold choice also interacts with spatial extent. ShakeAlert documents an initial restriction on how far alert regions extend, followed by an update using newer source information. The feature addresses uncertainty in very early estimates rather than claiming that distant places cease to be physically exposed during the restriction. The documented alert-pause feature.
No universally best threshold follows from the wave-speed calculation alone. An evaluation must state what counts as an unnecessary alert, what counts as a missed exceedance and how late forecasts are classified. Those definitions are part of the scientific result because they determine which outcomes are being counted.
An operational record is not one success percentage
The 2024 assessment of ShakeAlert's 2019–2023 operation illustrates the need to read classifications carefully. During the study period, the system created 95 events with maximum estimated magnitude at or above 4.5; 94 corresponded to real earthquakes. Seven met the project's internal false-event definition because of catalog-matching criteria, although all but one of those were poorly located real earthquakes. The study also identified missed or severely mislocated events near network boundaries. These figures describe source-event performance, not the fraction of people warned before damaging motion. Lux and colleagues' operational assessment.
A source-level error and a site-level error are different units of analysis. One poorly located earthquake can affect thousands of target locations. Conversely, an accurately identified event can produce a mixture of useful, late and unnecessary forecasts across its shaking region. Dividing successful detections by catalog earthquakes cannot establish a local warning success rate.
Consider an explicitly invented validation set with twenty earthquakes and one hundred instrumented target sites per earthquake. Suppose every earthquake is detected, but only forty event-site pairs exceed the chosen shaking threshold. Of those forty, twenty-four receive a timely forecast, eight receive a late one and eight receive none. Calling detection perfect would be correct for the first question and misleading for the second. The timely fraction among threshold exceedances is twenty-four out of forty, or sixty percent.
The other 1,960 pairs matter too, but not in the same denominator. Some may receive forecasts even though their observed motion stays below the threshold. Reporting only the total percentage of correctly classified pairs could make a system look impressive simply because threshold exceedances are rare. Reporting only the forty exceedances would hide the cost of forecasting them too broadly.
These invented numbers are not ShakeAlert performance estimates. They show why an evaluation needs the full contingency structure and a separate treatment of lateness. They also show why event-site observations are not automatically independent. A single source-location error can move many sites into the wrong category together.
Ground motion retains variability even with a known source
Improving source estimates cannot remove every unnecessary or missed local forecast. A 2019 study explicitly considered the remaining difficulty of forecasting ground-motion exceedances even with idealized knowledge of the earthquake source. It showed how variability in quantities such as acceleration and velocity changes the balance among alert outcomes. Its conclusions depend on the modeled criteria and user assumptions; they are not a license to describe all false alerts as unavoidable. Minson and colleagues' accuracy analysis.
An original comparison clarifies the remaining uncertainty. Suppose a model's central ground-motion estimate lies just below a chosen threshold. If actual motion varies around that estimate, some sites or events can exceed the threshold anyway. Raising the source magnitude estimate until every such exceedance is captured would also expand forecasts for cases that remain below it. A deterministic line on a map compresses this distribution into a binary boundary.
A useful investigation therefore asks whether uncertainty is calibrated across relevant conditions. Do nominally similar forecasts produce the expected range of outcomes? Are residuals systematically different near a network edge, in a particular basin or for a particular source style? Pooling all conditions can hide a repeatable bias inside apparently ordinary overall scatter.
The reference observations need scrutiny as well. Instruments sample selected locations; they do not measure every building or every patch of ground. If a verification map combines observations and modeled values, the evaluation should identify which are direct measurements and which depend partly on another model. Otherwise, apparent agreement can reflect shared assumptions rather than independent confirmation.
This does not make evaluation impossible. It makes the target more precise: assess the prediction against observations appropriate to its quantity, preserve uncertainty in the reference and avoid treating an uninstrumented location as if it had supplied a perfect measurement. The scientific improvement is often a better characterized limitation, not simply a higher headline score.
Replaying a large event without borrowing its future
Rare large earthquakes create a validation problem. Researchers cannot wait for a plentiful local sample of every rupture style before investigating performance, but simulated or replayed events are not equivalent to prospective operation.
The full 2025 version-three study describes tests aimed at larger earthquakes, while distinguishing them from real-time experience. Its production chronology also matters: version 3.0.1 went live in March 2024, before the paper's 2025 publication. The publication date is not the deployment date, and neither alone establishes the software version operating in every later test. The study's full manuscript.
A defensible replay needs a rule for what information the algorithm is permitted to know at each instant. Final magnitude, final rupture geometry and retrospectively corrected arrival picks cannot be injected into the early stages unless the experiment is explicitly estimating an idealized upper bound. Such a bound can answer a useful question, but it cannot be labeled achieved operational warning time.
Synthetic waveforms raise another set of questions. The simulated source might explore a rupture absent from the historical record, but it inherits assumptions about rupture timing, propagation and noise. A method tested on simulations closely resembling the models used to construct it may receive an easier examination than a method confronted with unanticipated real observations.
One original audit design would keep three columns separate: performance on untouched historical streams, performance on historical streams with controlled alterations and performance on fully synthetic scenarios. Each column would state the network configuration, software version, latency treatment and reference threshold. The same summary metric could then be compared without pretending the evidence has the same status.
Stress tests should include cases in which an earthquake grows differently from the expected template. They should also preserve periods of ordinary noise, because an estimator that looks excellent when told exactly where an earthquake begins may behave differently when required to find events in continuous operation. Testing the difficult rupture and testing the empty interval are complementary tasks.
A second route starts from shaking already observed
Not every research approach begins by estimating a source and then calculating its consequences. Ground-motion-based methods can use shaking already observed at stations to forecast conditions nearby. This changes the information pathway, although it does not abolish travel time or the need for adequate coverage.
PLUM was running on a ShakeAlert development server during the 2019 Ridgecrest sequence. A published analysis used that real-time record, including actual data availability and latency, and reported initial forecasts five and six seconds after the respective mainshock origins. The experiment evaluated several regional alerting strategies. It was development-system evidence, not a statement that PLUM was the production source of every public alert. The real-time PLUM study.
The conceptual contrast is useful. A source-based method can extrapolate from an estimated rupture to places not yet shaken strongly. An observation-propagation method can reduce dependence on some source assumptions but needs informative motion at suitable stations. Which route is earlier or more accurate depends on the event, network and target, not on the name of the method.
For a hypothetical comparison, place a strong-motion observation between a growing rupture and a target site. It may provide a valuable new constraint just when the source model remains uncertain. Move that station behind the rupture, or delay its data, and its usefulness can change sharply. A robust evaluation therefore compares methods on the same causal information stream rather than granting one an idealized input advantage.
The larger lesson is methodological diversity with explicit accounting. Combining methods can improve resilience when their weaknesses differ. It can also create false confidence if their errors share the same missing station, incorrect timing or reference model. Agreement is informative only after the dependencies behind it are understood.
Preserve the sequence, not just the final magnitude
A complete scientific record of one warning attempt should allow a reader to reconstruct the evolving opportunity. It needs the earthquake origin used for evaluation, the observations available to the algorithm, each consequential estimate and the local motion against which lead time is measured. The exact record format can vary; the causal order cannot.
The official post-event summary page itself draws an important boundary: its estimated warning times exclude the additional interval used by licensed operators to deliver alerts to end users. It also distinguishes automatically generated summaries from the subset receiving human review. A readily accessible summary is therefore a starting point for analysis, not a substitute for understanding its timing and review status. ShakeAlert's explanation of post-event summaries.
An original diagnostic ledger could record, for each target and threshold, the first qualifying prediction, subsequent revisions, the observed threshold-crossing time and any interval when the required observations were unavailable. Negative lead times should remain negative. Excluding them from a distribution of warning times would silently remove some of the most important failures.
The ledger should also distinguish an estimate that was never produced from one that was produced and later withdrawn. Those are different scientific histories. So are a location correction, a growing magnitude and a revised local-shaking calculation. A final event page can conceal all three if only its most recent values survive.
Earthquake early warning is strongest when its claims are this specific. It can sometimes turn a physical separation between observation and consequence into useful seconds. It cannot promise the same interval everywhere, infer every future rupture from its opening motion or transform a completed archive into proof of timely knowledge. The meaningful achievement is not that a machine said earthquake quickly. It is that, for a defined place and level of shaking, the available observations supported a sufficiently informative forecast while time still remained.
<!-- ARTICLE_BODY_END -->