{"id":"c454a253-2e11-4b5e-b541-0e69e1eb4247","arxiv_id":"2607.10450","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"low","formal_verification":"none","parameter_count":4,"one_line_summary":"VORA associates earthquake phases via unsupervised spatio-temporal clustering of origin times on a Voronoi station graph, delivering fastest runtime and local-to-global scalability without training.","lead":"VORA groups seismic P and S arrivals into earthquakes by clustering estimated origin times across Voronoi-neighbor stations, without training. It is the fastest of three tested associators and runs unchanged from local arrays to a global station set, which matters as networks densify and automated picks explode.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"Synthetic stress-test recall/precision and runtime claims rest on an idealized dual-phase assumption that is not stress-tested under realistic missing-S rates.","rationale":"The Reader correctly flags the dual-phase premise as the weakest assumption and already notes the mainshock miss. The present stress-test simply sharpens that observation: the quantitative support for the strongest claim (Fig. 5) is generated under the very idealization that the paper later acknowledges fails in practice. Because the real-data experiments do not isolate missing-S rates, they cannot substitute for a controlled ablation. The concern does not overturn the method’s conceptual novelty or its demonstrated speed under ideal conditions, so the verdict remains CONDITIONAL (pending code release and the missing-S quantification). No stronger internal inconsistency or circularity appears; the Voronoi construction and origin-time clustering are clearly described and free of derivation gaps. Agreement with the Reader is therefore full on the identity of the load-bearing soft spot.","tokens_in":15971,"tokens_out":669,"duration_ms":11447,"concrete_test":"Re-run the eight synthetic sequences of §3.1 after randomly dropping S picks at rates of 20 %, 40 % and 60 % (matching observed PhaseNet S-miss rates near the Ridgecrest mainshock). Recompute event-level recall/precision and wall-clock time for VORA, GaMMA and PyOcto under identical hardware. If VORA’s recall at ∆t̄ = 5 s falls below ~0.7 or loses its order-of-magnitude speed advantage once S completeness drops below ~70 %, the headline claim weakens under realistic conditions.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The strongest claim (fastest runtime + recall >0.8 / precision >0.9 at ∆t̄ = 5 s, same hyper-parameters local-to-global) is supported by the synthetic suite in §3.1 and Fig. 5. That suite, however, explicitly assumes every station always records both P and S (\"We assume that each station always detects both P and S phases from an earthquake\"), so Eq. 1 always yields a usable origin-time interval and every station can participate in the Voronoi-neighbor graph. The paper itself shows that real pick streams violate this: the Ridgecrest mainshock and other events are missed precisely because of missing S picks (Discussion, Fig. 9). Consequently the reported high-recall / high-precision numbers under intense seismicity, and the associated runtime advantage, have not been demonstrated under the dual-phase incompleteness that actually occurs. The real-data recall of 0.86 (local Ridgecrest) and the global-scale daily recalls are measured against catalogs, not against ground-truth associations under controlled missing-S rates, so they do not close the gap. The optional EM sub-clustering step cannot recover events that never formed a dual-phase seed. This is the single most load-bearing soft spot for the central performance claim.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The manuscript introduces VORA, an unsupervised earthquake phase associator that estimates hypothetical origin times from P–S pairs (Eq. 1, or direct waveform prediction) and clusters them over Voronoi-defined station neighborhoods, with an optional EM sub-clustering step (Eq. 2) for events that merge in space–time. The central claims are that VORA is training-free, governed by few physically interpretable hyperparameters, achieves the fastest runtime among the three methods tested, maintains high recall and precision under dense seismicity, and uses the same framework from a local 121-station network to a 4461-station global pick set covering the 2019 Ridgecrest sequence. Support comes from eight synthetic sequences (5000 events each, average inter-event times from 64 s to 5 s), a local Ridgecrest PhaseNet pick set compared with GaMMA and PyOcto, and a two-week global pick subset with USGS/SCSN catalog matching.","tokens_in":16287,"tokens_out":1422,"duration_ms":36057,"significance":"If the performance and scalability claims hold under realistic pick incompleteness, VORA would be a practically important contribution: a fast, training-free associator whose station graph adapts to non-uniform and evolving networks without region-specific retraining or absolute velocity models. The Voronoi neighborhood construction is a clear methodological strength relative to fixed-K or pure distance graphs, and the demonstrated local-to-global reuse of the same hyperparameter set is rare among published associators. The runtime advantage on a single CPU (Fig. 5) is valuable for large automated pick streams. These strengths are real even if the intense-seismicity numbers need more careful qualification.","major_comments":[{"comment":"§3.1 and Fig. 5: The synthetic stress test that underpins the strongest performance claim (recall >0.8 and precision >0.9 at Δt̄ = 5 s, plus the associated runtime ranking) explicitly assumes that every station always records both P and S. The manuscript states this assumption directly and does not inject controlled missing-S or false-pick rates. Because Eq. 1 and the dual-phase seed construction require usable P–S pairs, those headline numbers have not been demonstrated under the incompleteness that the Discussion and Fig. 9 show is decisive on real data (including the missed Ridgecrest mainshock). A controlled missing-S / missing-P sweep on the same synthetic suite, or a clear separation of claims between dual-phase-complete and realistic pick streams, is needed before the intense-seismicity claim can stand as written.","section":"§3.1 Synthetic Stress Test; Fig. 5"},{"comment":"§3.2–3.3 and Abstract: Real-data recall (0.86 local Ridgecrest; daily global recalls in Fig. 8) is measured by matching associated events to SCSN/USGS catalogs, not by pick-level association accuracy under known ground-truth associations with realistic dual-phase incompleteness. Catalog matching is standard but does not close the gap left by the idealized synthetics: events that never form dual-phase seeds cannot be recovered by the optional EM step. The Abstract’s phrasing that VORA “maintains robust performance (high recall and precision) even under intense seismicity” currently blends the dual-phase-complete synthetic result with real-data catalog recall and should be tightened accordingly.","section":"Abstract; §3.2; §3.3; Fig. 8–9"},{"comment":"§2.4, Eq. (2): The optional EM sub-clustering relies on a constant pseudo-slowness K per event and seed stations identified as local minima of S–P intervals. The paper notes that quality “depends heavily on how precisely the seed stations are identified” and that the approach works better for denser networks, but provides no quantitative success rate for the sub-clustering step across the synthetic suite (e.g., fraction of merged clusters correctly split, sensitivity to seed errors). Because this step is the only remedy offered for the hardest overlapping cases that drive the Δt̄ = 5 s regime, its failure modes should be reported with the same recall/precision protocol used for the main clustering pass.","section":"§2.4 Optional EM-Based Clustering; Eq. (2); Fig. 3–4"}],"minor_comments":[{"comment":"Clarify in §2.1–2.3 whether origin-time intervals from a VP/VS range are treated as closed intervals for overlap, and how multi-ratio candidates from the same station are deduplicated before DBSCAN-style growth.","section":"§2.1–2.3"},{"comment":"Fig. 5d: state explicitly whether runtime includes only association or also preliminary location, and whether GaMMA/PyOcto were run with their recommended parallel settings disabled solely for fairness.","section":"Fig. 5"},{"comment":"§3.3: the 20 s maximum P–S separation (local-only detection) is important; state it earlier when claiming “global-scale” association so readers do not infer teleseismic association capability.","section":"§3.3"},{"comment":"Fig. 6: VORA’s systematically larger epicentral residuals are attributed to requiring both P and S; a short quantitative note on average station count per event versus GaMMA/PyOcto would make that trade-off clearer.","section":"Fig. 6; §3.2"},{"comment":"Minor wording: “archiving comparable” → “achieving comparable” (Introduction); ensure consistent spelling of Munchmeyer/Münchmeyer across text and references.","section":"Introduction; References"},{"comment":"Open Research: code is promised after peer review; a frozen commit or archive at revision would strengthen reproducibility claims for a methods paper.","section":"Open Research Section"}],"recommendation":"major_revision","confidential_remarks":"The dual-phase incompleteness issue is the only load-bearing soft spot; the method and Voronoi framing are otherwise solid and appropriate for a methods journal. I would not reject on novelty grounds. If the authors add a missing-phase synthetic sweep (or clearly demote the dual-phase-complete numbers), this could move to minor revision or accept quickly. Fit for physics.geo-ph / seismology methods venues is good."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The one thing worth knowing: VORA is a clean, training-free phase associator that clusters origin-time intervals on a Voronoi station graph and is an order of magnitude faster than GaMMA and PyOcto on the same hardware, with the same hyper-parameters working from a 121-station local net to a 4461-station global pick set.\n\nWhat is actually new is the combination. Origin-time estimates from P–S pairs (or direct waveform prediction) and Voronoi neighborhoods each exist in the literature, but packaging them into an unsupervised spatio-temporal associator that deliberately skips absolute velocity models and grid search, plus the higher-order neighbor construction by successive removal, is a finished method that was not already on the shelf. The optional EM sub-clustering for near-coincident events is a sensible add-on. The synthetic suite (eight sequences of 5000 events, average inter-event times down to 5 s) uses clear matching (Hungarian + 0.5 pick-overlap) and reports both recall/precision and wall-clock time; real Ridgecrest comparisons against SCSN and the two external methods give ~0.86 recall, and the global two-week run recovers the expected conjugate geometry and daily USGS recall patterns. Free parameters are few and physically interpretable. Citation pattern is normal and the math is elementary but correctly applied.\n\nThe soft spot the stress-test flags is real and load-bearing for the strongest claim: the synthetic stress tests assume every station always records both P and S, so Eq. 1 always yields a usable interval. Real streams do not; the paper itself shows the Ridgecrest mainshock and other events are missed for lack of S picks (Discussion, Fig. 9). Real-data recalls against catalogs do not fully close that gap, and the EM step cannot invent dual-phase seeds that never formed. That said, the paper is transparent about the premise, the runtime advantage is still real under the conditions tested, and the local-to-global scaling with fixed hyper-parameters is demonstrated. Code is promised post-review, so reproducibility is currently moderate.\n\nThis is for people who build or run catalogs and need something fast that does not retrain when the network or region changes. It deserves a serious referee. I would engage with it and expect it to be useful once the dual-phase sensitivity is quantified more carefully.","headline":"Solid, practical associator that is genuinely fast and scale-agnostic; the dual-phase premise is a real limit but the paper owns it and the rest of the evidence holds.","tokens_in":16894,"tokens_out":574,"would_cite":true,"duration_ms":8319,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":["91.30.Ab","91.30.Dk","93.85.Bc"],"model":"grok-4.5","headline":"VORA turns earthquake phase association into unsupervised clustering of origin times on a Voronoi station graph, and the same code runs from local arrays to a global pick set at record speed.","keywords":["earthquake phase association","Voronoi tessellation","origin-time clustering","unsupervised clustering","seismic network geometry","Ridgecrest sequence","global-scale seismicity","DBSCAN"],"falsifier":"Run the identical VORA configuration on a dense aftershock sequence whose S-phase detection rate is deliberately degraded below the level of the Ridgecrest mainshock window; if recall collapses while competing methods that use P-only or multi-phase logic remain stable, the central claim fails.","tokens_in":16856,"feed_emoji":"🌍","tokens_out":625,"duration_ms":9465,"temperature":0.7,"pith_summary":"Modern seismic networks and deep-learning pickers flood analysts with millions of P and S arrivals that must be grouped into common earthquake origins. VORA estimates a candidate origin time from each P–S pair (or from a waveform model) using only a narrow range of velocity ratios, then asks which of those times come from neighboring stations. Neighbors are defined by the Voronoi diagram of the operating stations, so the spatial graph automatically adapts when stations are added or removed and works unchanged from a dense local array to a global network. A modified DBSCAN clusters the origin-time intervals across those neighbors; an optional EM step splits merged clusters when two events fall inside the same cells. On synthetic catalogs with average inter-event times as short as 5 s, and on real Ridgecrest picks at both local and global scales, VORA is the fastest of the three methods tested and keeps high recall and precision with a single hyper-parameter set. The result is a training-free associator that can keep pace with the growing volume of automated picks.","feed_headline":"VORA clusters earthquake picks by origin time on a Voronoi graph","feed_subtitle":"Same code runs local-to-global, finishes fastest under intense seismicity, needs no training","key_machinery":"Voronoi station graph plus origin-time intervals: each station owns a Voronoi cell; stations that share an edge are neighbors; candidate origin times (Eq. 1 or waveform prediction) are clustered only across those neighbors, turning association into a single unsupervised pass.","core_discovery":"The paper claims that earthquake phase association reduces to unsupervised spatio-temporal clustering of hypothetical origin times once station neighborhoods are defined by Voronoi tessellation, and that this formulation is both faster and more scale-invariant than existing grid-search or trained methods.","pith_inferences":[],"forward_implications":[],"fun_headline_variants":["VORA associates quake phases via origin-time clusters on Voronoi neighbors","Voronoi station graph plus origin times enables rapid phase association","Unsupervised origin-time clustering with Voronoi adjacency links seismic phases","VORA scales phase association local-to-global without training or grids","Origin-time clusters on Voronoi edges group earthquake phases fastest"],"cache_read_input_tokens":128,"weakest_assumption_plain":"Enough stations must report both a usable P and an S pick (or a direct origin-time estimate) so that neighboring origin-time intervals can form a cluster; missing S phases cause real events, including the Ridgecrest mainshock, to be lost.","fun_headline_variants_meta":{"raw":{"variants":["VORA associates quake phases via origin-time clusters on Voronoi neighbors","Voronoi station graph plus origin times enables rapid phase association","Unsupervised origin-time clustering with Voronoi adjacency links seismic phases","VORA scales phase association local-to-global without training or grids","Origin-time clusters on Voronoi edges group earthquake phases fastest"]},"model":"grok-4.5","effort":"low","cost_usd":0.004922,"raw_usage":{"total_tokens":1423,"prompt_tokens":808,"num_sources_used":0,"completion_tokens":93,"cost_in_usd_ticks":49220000,"prompt_tokens_details":{"text_tokens":808,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":522,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":808,"tokens_out":93,"duration_ms":7566,"temperature":1.0,"reasoning_tokens":522,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-14T11:36:45.715340+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Run the identical VORA configuration on a dense aftershock sequence whose S-phase detection rate is deliberately degraded below the level of the Ridgecrest mainshock window; if recall collapses while competing methods that use P-only or multi-phase logic remain stable, the central claim fails.","supporting_citations":[],"review_version":1}