{"id":"bb1e0d5d-3783-450c-b4f9-e071ffc0924b","arxiv_id":"2607.08975","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"AEA detected the Mw 6.2 Marmara Ereğlisi earthquake in 5.31 s, alerted 16 M users with 90% TP / 99% precision and median 56 s warning for MMI III, and elicited high rates of protective actions and trust when alerts arrived before shaking.","lead":"Google's Android Earthquake Alert system detected the 2025 Mw 6.2 Marmara Ereğlisi quake in 5.31 seconds and warned over 16 million users with a median 56-second lead time for weak shaking at 90% true-positive rate. The case study shows dense phone networks can outpace traditional stations near coasts and that timely alerts drive protective actions and user trust.","discovery_kind":"extension","skeptic_critique":{"model":"grok-4.5","headline":"No significant objection identified beyond the reader's already-flagged ground-truth contour assumption.","rationale":"The Reader's weakest_assumption is precisely the load-bearing point: the composite ground-truth contour that defines the denominator for the 90 % true-positive and 99 % precision claims. I re-checked the Methods section, Figure S3, and the spatial-grid classification rules and found no stronger technical vulnerability (e.g., no circular use of AEA triggers in the ground truth, no unstated magnitude-saturation correction that would reverse the Take-Action under-coverage, no evidence that the 5.31 s detection time is an artifact of clock skew). The behavioral results are correlational and self-selected, yet they are presented with effect sizes and are secondary to the technical performance numbers. Because the Reader already conditioned the verdict on exactly this contour uncertainty (plus proprietary data and the M6.2 sweet-spot character), no further downward adjustment is warranted. The concrete test above would settle residual doubt about contour sensitivity without requiring proprietary phone data.","tokens_in":30662,"tokens_out":585,"duration_ms":7637,"concrete_test":"Recompute the population-weighted TP/FN rates of Figure 4 after replacing the composite MMI-III contour with the pure USGS ShakeMap MMI-III contour alone (and, separately, with the pure INGV contour). If either pure-contour TP rate falls below ~80 % or precision below ~95 %, the headline 90 %/99 % figures become contour-dependent and the claim weakens; otherwise the multi-source construction is robust.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central technical claim (5.31 s detection, >16 M alerts, 90 % TP / 99 % precision, median 56 s warning for MMI-III Be Aware) rests on a composite ground-truth contour assembled from USGS/INGV ShakeMaps, felt reports, and Google Search survey yes-ratios (Methods; Fig. S3). The Reader correctly identifies this as the weakest assumption. After re-examining the construction, the contour is multi-sourced, the spatial-grid metrics are population-weighted rather than device-weighted, and the paper itself notes the sampling bias of the Search survey. No additional internal inconsistency, hidden selection effect, or unacknowledged circularity appears that would further undermine the reported numbers. Behavioral associations (timeliness \to usefulness/trust/active response) are supported by large-N chi-square tests with reported effect sizes (Table 1) and are consistent with the independent (though small) DYFI subsample. Proprietary phone data and self-selection remain real but already-cited limitations that do not invalidate the observational audit of this single event.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"This manuscript evaluates Google’s Android Earthquake Alert (AEA) system for the 23 April 2025 Mw 6.2 Marmara Ereğlisi earthquake. It reports detection 5.31 s after origin time, issuance of alerts to >16 million users, 90 % true-positive rate and 99 % precision for MMI-III “Be Aware” alerts (median 56 s warning time, up to ~150 s), faster phone-network detection than nearby AFAD stations despite the offshore epicenter, and large-N user-feedback analyses linking pre-shaking alert arrival to higher rates of protective action (DCHO or warning others), perceived usefulness, and future trust. Performance is quantified on population-weighted ~20 km grids against a multi-source MMI-III ground-truth contour; behavioral associations are tested with chi-square statistics and Cramer’s V effect sizes (Table 1).","tokens_in":30954,"tokens_out":1136,"duration_ms":20013,"significance":"If the reported metrics hold, the paper supplies a rare, large-scale observational audit of a production smartphone-based EEW system under real seismic loading, combining technical performance (detection latency, magnitude evolution, warning-time distributions) with behavioral response data at N~67 000. The multi-source ground-truth construction, population-weighted TP/LA/FN/FP definitions, explicit effect sizes, and open station-processing code are clear strengths. The work is of direct interest to seismologists, EEW operators, and risk-reduction communities because it quantifies both the technical reach of crowd-sourced sensing and the conditions under which alerts translate into protective action and trust.","major_comments":[{"comment":"Methods (“Alert performance evaluation”) and Fig. S3: the composite MMI-III ground-truth contour that underpins the 90 % TP / 99 % precision claim mixes USGS/INGV ShakeMaps, DYFI/EMSC felt reports, and Google Search “yes” ratios. The Search survey samples only users who actively sought earthquake information and therefore already perceived shaking; the paper itself notes this selection. Because the outer edges of the TP region (e.g., Izmir) rest partly on these ratios, a sensitivity test that recomputes the spatial metrics after excluding the Search layer is needed to confirm that the headline rates are not inflated.","section":"Methods, Alert performance evaluation; Fig. S3"},{"comment":"User Feedback Analysis and Table 1: the chi-square associations between alert timing, shaking intensity, active response, usefulness, and future trust rest on self-selected in-alert respondents (N=67 056) plus a tiny independent DYFI subsample (N=55). While the paper correctly flags non-random sampling, the claim that AEA “elicited substantially higher rates of protective action” relative to prior studies (Goltz, Nakayachi, Vinnell) requires a more quantitative discussion of how self-selection and the recent 2023 Kahramanmaraş sequence may bias the 46 % active-response figure upward.","section":"User Feedback Analysis; Table 1; Discussion"}],"minor_comments":[{"comment":"Discussion paragraph comparing prior studies: “greater percentage of infections among responders” is a clear typographical error for “inactions.”","section":"Discussion"},{"comment":"Abstract and title: “Googles” and “Marmara Ereglisi” lack the correct apostrophe/diacritics used elsewhere; standardize to “Google’s” and “Marmara Ereğlisi.”","section":"Abstract; title"},{"comment":"Figure 1 caption and text: the claim that phones detected the event “sooner than conventional seismic stations” is supported for the nearest coastal phones, but the two AFAD stations at 25–29 km are the only stations inside 30 km; a short quantitative comparison of first-trigger times (phones vs. stations) would make the statement more precise.","section":"Results, Detection and Strong Motion Arrivals; Fig. 1"},{"comment":"Table S1 lists successive magnitude estimates and alert radii; adding the corresponding alert-issue times relative to origin would help readers reconstruct the timeline shown in Fig. 3 without cross-referencing the text.","section":"Supplementary Table S1"},{"comment":"Ethics / Data availability: the statement that AEA phone-trigger and feedback data “cannot be made publicly available” is understandable, yet the paper would benefit from a brief note on whether any aggregated, privacy-preserving summary tables (e.g., binned warning-time histograms) can be released to support independent verification of the population-weighted metrics.","section":"Data availability"}],"recommendation":"minor_revision","confidential_remarks":"The majority of authors are Google employees and the work is funded by Google LLC; competing-interest disclosures are present and appear adequate. The proprietary nature of the phone data is a real but already-acknowledged limitation for full reproducibility; the open station-processing code and multi-agency ground-truth sources mitigate this for the technical performance claims. Scope is appropriate for a geophysics / EEW journal."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"This is the first detailed public performance-plus-behavior audit of Google’s AEA system for a moderate-to-large event. Detection at 5.31 s after origin, >16 M alerts, population-weighted 90 % true-positive / 99 % precision for MMI-III “Be Aware” with median 56 s warning, and the large-N survey results (N ≈ 67 k) linking pre-shaking arrival to active protective response, usefulness, and future trust are new empirical numbers that were not in Allen et al. 2025 or the smaller-event behavioral papers.\n\nWhat they did well: external ground truth (USGS ShakeMap, AFAD stations, EMSC/USGS felt reports), transparent TP/LA/FN/FP definitions on population grids, magnitude-evolution and warning-time distributions, and chi-square tests with Cramer’s V effect sizes (Table 1). Phone waveforms show denser coastal triggers outpacing the sparse near-source AFAD stations for this offshore epicenter—an honest demonstration of the density advantage without overclaiming global superiority. The discussion correctly flags the M6.2 “sweet spot,” aftershock difficulties, and the Be Aware vs Take Action distinction (magnitude saturation left most MMI-V inland users without the stronger alert).\n\nSoft spots are real but already acknowledged and proportionate. The composite MMI-III contour (ShakeMaps + felt reports + Google Search “yes” ratios) is the weakest assumption; the Search survey is self-selected and samples motivated feelers, so distant TP rates could be mildly inflated. Proprietary phone data block full independent replication. Surveys are voluntary. None of these invent circularity or free parameters; the technical metrics remain observational against external catalogs.\n\nThis is for EEW operators, disaster-resilience researchers, and anyone who needs real-event numbers rather than simulations. It deserves a serious referee. I would engage with it, cite the detection latency, population metrics, and behavioral associations, and bring it to reading group.","headline":"Solid large-N observational audit of AEA on a real M6.2: 5.31 s detection, 90 % TP / 99 % precision for Be Aware, median 56 s warning, and clear lead-time → action/trust links; soft spots are the composite MMI-III contour and proprietary data, not fatal.","tokens_in":31555,"tokens_out":550,"would_cite":true,"duration_ms":7632,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"Android phones detected the 2025 Marmara quake in 5.31 seconds and gave 16 million users a median 56-second warning for weak shaking.","keywords":["earthquake early warning","Android Earthquake Alert","smartphone seismology","crowd-sourced sensing","warning time","user response","Marmara Sea","MMI intensity"],"falsifier":"An independent, instrument-based intensity map that places a substantially different fraction of the regional population outside the AEA “Be Aware” polygons would drop the true-positive rate well below 90 percent and falsify the claimed precision.","tokens_in":31570,"feed_emoji":"📱","tokens_out":996,"duration_ms":11636,"temperature":0.7,"pith_summary":"This paper evaluates how Google’s Android Earthquake Alert system performed during the Mw 6.2 Marmara Ereğlisi earthquake. The phone network detected the event 5.31 seconds after origin, faster than the sparse local seismic stations near the offshore epicenter, and issued alerts to more than 16 million users. For weak (MMI III) shaking the system delivered a median 56-second warning (up to 150 seconds), with 90 percent true-positive coverage of the population and 99 percent precision. User surveys show that people who received the alert before shaking were far more likely to take protective action, rate the alert useful, and say they would trust future alerts. The authors argue that dense, crowd-sourced smartphone sensing can both speed detection and, when the alert is timely, drive constructive public response at continental scale.","feed_headline":"Phones spotted the Marmara quake in 5.31 s, warned 16 million","feed_subtitle":"Median 56-second lead time and high protective-action rates show crowd-sourced EEW can work at scale","key_machinery":"The AEA detection-and-alert pipeline: on-phone accelerometers that flag P- or S-wave arrivals, a backend that fuses the resulting sparse triggers into real-time magnitude and intensity polygons, and two-tier delivery (“Be Aware” for MMI III–IV, “Take Action” for MMI ≥ V) whose spatial coverage is scored against a composite ground-truth contour using population-weighted true-positive / false-negative metrics.","core_discovery":"During the Mw 6.2 Marmara Ereğlisi earthquake the Android Earthquake Alert system detected the event 5.31 s after origin, alerted more than 16 million users, achieved a 90 percent true-positive rate and 99 percent precision for MMI-III “Be Aware” alerts with a median 56 s warning, and produced substantially higher rates of protective action, perceived usefulness and future trust when the alert arrived before shaking.","pith_inferences":["Because magnitude saturation limited the final “Take Action” polygon, the paper’s high true-positive figure is driven almost entirely by the lower-tier “Be Aware” alerts; a larger event would expose a more severe performance gap.","The elevated active-response rate relative to earlier studies in other countries may reflect both the recent 2023 Kahramanmaraş disaster and cultural norms rather than a universal property of the AEA interface.","Self-selected survey respondents who actively searched for earthquake information may over-represent people who already felt shaking, inflating the apparent correlation between alert receipt and protective behavior.","If phone density continues to grow faster than traditional station density, the relative detection advantage of crowd-sourced networks will widen for most inhabited coastal and inland settings."],"forward_implications":["In dense coastal or urban corridors, phone networks can issue usable early warnings even for offshore events that traditional stations detect later.","Alert-before-shaking is the dominant driver of protective action, usefulness ratings and future trust; late alerts actively erode credibility.","Hybrid phone-plus-station architectures could reduce detection latency and improve magnitude estimates for complex aftershock sequences.","“Take Action” breakthrough alerts remain essential for nighttime or Do-Not-Disturb scenarios where standard notifications fail to wake users.","Public-education campaigns that reinforce Drop-Cover-Hold-On remain necessary so that alerts trigger prescribed actions rather than panic."],"fun_headline_variants":["Android spotted Marmara quake in 5.31s and alerted 16M phones","Phones gave 16M users median 56s warning for Marmara quake","Android EEW hit 90% true positives alerting 16M in Marmara event","Crowd phones beat stations, warned millions before Marmara shaking","Pre-shake Android alerts drove protective action and trust at scale"],"cache_read_input_tokens":16512,"weakest_assumption_plain":"The composite MMI-III contour built from ShakeMaps, felt reports and Google Search “yes” ratios correctly defines the population that should have been alerted, so that the reported true-positive and false-negative rates are unbiased.","fun_headline_variants_meta":{"raw":{"variants":["Android spotted Marmara quake in 5.31s and alerted 16M phones","Phones gave 16M users median 56s warning for Marmara quake","Android EEW hit 90% true positives alerting 16M in Marmara event","Crowd phones beat stations, warned millions before Marmara shaking","Pre-shake Android alerts drove protective action and trust at scale"]},"model":"grok-4.5","effort":"low","cost_usd":0.006422,"raw_usage":{"total_tokens":1633,"prompt_tokens":748,"num_sources_used":0,"completion_tokens":104,"cost_in_usd_ticks":64220000,"prompt_tokens_details":{"text_tokens":748,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":781,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":748,"tokens_out":104,"duration_ms":5701,"temperature":1.0,"reasoning_tokens":781,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-13T01:16:56.440718+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"An independent, instrument-based intensity map that places a substantially different fraction of the regional population outside the AEA “Be Aware” polygons would drop the true-positive rate well below 90 percent and falsify the claimed precision.","supporting_citations":[],"review_version":1}