{"id":"08018a91-6576-4349-becd-8f0bcd8acb66","arxiv_id":"2608.09804","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"SGNAX unifies snax and omicron in one Python-native pipeline and reproduces production trigger populations while exposing a low-frequency SNR amplitude error in snax.","lead":"SGNAX is a new open-source Python pipeline that combines two standard gravitational-wave glitch-finding methods, matched filtering and Q-transform excess power, in one streaming dataflow graph. The authors show it reproduces existing production triggers to within a few percent in the key 10-100 Hz band and uncover a square-root-of-two SNR calibration bug in the production snax tool.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 1/sqrt(2) production-snax amplitude error in the 12.8-25.6 Hz band is inferred from a cross-pipeline SNR ratio; a direct injection through the production snax chain is needed to rule out a scale error in sgnax's own low-rate path.","rationale":"The paper makes three intertwined claims: a unified pipeline with competitive performance, reproduction of the omicron trigger population, and discovery of a production snax amplitude bug. The first two are well supported: injection recovery is 99.4% on both modes, the Q-transform runtime is directly comparable to omicron, multi-channel amortization is demonstrated, and the 80% per-bin coincidence on production auxiliary channels is transparently presented. The production-snax amplitude-error claim is the weakest link because it is an indirect inference from a ratio between two independently implemented chains. The exact 1/sqrt(2) value and the octave-boundary step are suggestive, but they cannot by themselves identify which chain carries the scale error; the paper's own static audit did not localize the excess, and the low-frequency calibration of sgnax is asserted rather than shown. A direct injection through the production snax chain would settle the attribution. If the proposed test falsifies the production-bug claim, the paper's main unified-pipeline contribution still stands, but the amplitude-error finding would need to be retracted or substantially weakened. The reader's conditional verdict already reflects this uncertainty, so no verdict change is needed.","tokens_in":29466,"tokens_out":6432,"duration_ms":60196,"concrete_test":"Run the production snax executable (reference [18]) on a synthetic unit-variance Gaussian channel with sine-Gaussian injections of analytically known optimal SNR at f = 20 Hz, Q = 15, and also at f = 10 Hz, using the production tapered bank and the same 12-100 Hz configuration. If the recovered production SNR matches the analytic value to a few percent, the 0.705 ratio reflects a scale error in sgnax's low-frequency chain; if production reports approximately sqrt(2) times the analytic SNR in the 12.8-25.6 Hz band, the production amplitude error is confirmed. Repeat through the full production feature path, not just the template normalization.","verdict_should_be":"UNCHANGED","load_bearing_attack":"In Sec. VH, the conclusion that production snax over-amplifies its 12.8-25.6 Hz band by sqrt(2) is derived almost entirely from a cross-pipeline ratio: sgnax-timeseries recovers 0.70 of the production SNR on coincident loud features, the ratio steps at the nominal 64 Hz rate-band boundary, and sgnax's own chain is claimed to be injection-calibrated down to 12.7 Hz. The inference is sound only if the two chains differ by precisely the production band gain. But the paper states that a static audit of the production chain did not localize the excess, and the low-frequency calibration of sgnax (injections at 12.7 Hz through the full path) is reported in text without a figure or table. A constant 0.705 ratio is equally consistent with an under-amplification by sqrt(2) in sgnax's low-rate resampler or whitener, or with a factor-of-two effective template-normalization difference (e.g., tapered versus untapered templates), because any band-structured gain in either pipeline produces the same boundary-stepped signature. The load-bearing step is therefore the attribution of the entire ratio to production snax, not any internal inconsistency in the population-level comparisons.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper describes SGNAX, a Python-native pipeline built on the Stream Graph Navigator framework, which unifies the multi-rate sine-Gaussian matched filter of snax and the Q-transform excess-power search of omicron in a single dataflow graph. It claims a shared configuration object and output schema, multiple data-source backends for offline and online use, multi-channel processing with decreasing per-channel cost, matched-filter end-to-end latency of about five seconds, and a full day of 16 kHz strain analyzed in 14 minutes by the Q-transform mode. The validation includes injection-recovery tests for both modes, false-alarm/efficiency curves, comparisons with production omicron and snax outputs on 24 hours of LIGO data, runtime and latency benchmarks, and a streaming glitch-rate frame product. The most consequential claim is that the injection-calibrated reimplementation reveals a multiband amplitude error in production snax that inflates reported SNRs below 25.6 Hz by factors of sqrt(2) to 2, with the mechanism identified as incorrect octave-rate-band amplitude compensation.","tokens_in":29748,"tokens_out":4662,"duration_ms":70475,"significance":"If substantiated, SGNAX would be a valuable consolidation of two heavily used detector-characterization tools, and the open-source release with its test suite is a genuine reproducibility contribution. The injection-recovery studies, the transparent omicron comparison, the false-alarm/efficiency characterization, and the measured runtime and latency data are concrete strengths that go beyond a purely architectural paper. However, the headline few-percent agreement with omicron is restricted to a band selected after inspection, and the production snax amplitude-error claim depends on attributing a cross-pipeline SNR ratio to a production-side bug without a direct injection through the production chain. These load-bearing claims need additional support before the paper's central conclusions can be accepted at face value.","major_comments":[{"comment":"The conclusion that production snax over-amplifies its 12.8-25.6 Hz band by sqrt(2) rests on the measured median SNR ratio of 0.705 between sgnax-timeseries and production snax on coincident loud features, together with the assertion that sgnax's own whitening and normalization chain has the textbook matched-filter scale of Eqs. (1)-(3). That assertion is reported in text without a figure or table for the low-frequency calibration, and the paper explicitly states that a static audit of the production chain did not localize the excess. A constant 0.705 ratio is equally consistent with an under-amplification by sqrt(2) in sgnax's low-rate resampling or whitening path, or with a factor-of-two template-normalization difference, because any band-structured gain in either pipeline produces the same boundary-stepped signature. The load-bearing step is the attribution of the entire ratio to production snax, so the authors should add a direct injection test through the production snax chain, or an equivalent independent calibration of both pipelines, before presenting the production bug as established.","section":"Sec. VH, Fig. 12(b)"},{"comment":"The abstract and conclusion state that SGNAX reproduces the omicron trigger population to a few percent, but the few-percent metrics in Table IV are computed in the 10-100 Hz, SNR>8 window that was identified after a frequency-SNR sweep over the data, and the full-band comparison shows a 12% total deficit concentrated in near-threshold high-frequency tiles whose origin is left to future work. The few-percent claim should be explicitly scoped to that band, and the band-selection process should be described as exploratory rather than as the predetermined operational window. The high-frequency residual should either be included in the headline metrics or quantified separately in the same table, with a statement of whether it affects the detector-characterization conclusions for which the omicron comparison is offered as evidence.","section":"Sec. VF/VG and Table IV"},{"comment":"The snax column of Table IV reports a trigger-rate ratio of 0.46 and an SNR-distribution KS statistic of 0.49, and the text attributes all of these residuals to the production amplitude scale. The per-channel breakdown of the 80% per-bin coincidence and the number of coincident loud features underlying the SNR and time medians should be reported, and the rate/KS metrics should be recomputed after applying the proposed 1/sqrt(2) correction to the production SNRs in the affected band. Without that demonstration, the statement that threshold- and rank-based consumers of the production features are unaffected by the scale error is not directly supported by the table.","section":"Sec. VH and Table IV, snax column"}],"minor_comments":[{"comment":"The typesetting has many missing spaces and ligature artifacts, for example 'We presentsgnax', 'a singlepython', 'gra vity spy', and 'a median of0.43'; a careful proofreading pass is needed before publication.","section":"Throughout"},{"comment":"The abstract says that a full day of 16 kHz strain is analyzed in 14 minutes, but Table V shows that this is the Q-transform figure; the matched-filter mode takes 84 minutes. Please qualify the abstract statement accordingly.","section":"Abstract and Table V"},{"comment":"The coincidence pairing for the omicron comparison uses a 0.2 s window and a factor-of-two frequency tolerance, but only the fraction of omicron triggers that find a partner is reported. The fraction of sgnax triggers that find an omicron partner should also be reported, since asymmetric pairing can bias the median per-trigger SNR difference.","section":"Sec. VG"},{"comment":"The bottom panel shows matched-filter SNR for sgnax-extract while the top two panels show excess-power SNR, and the caption notes that the shared color scale is only a morphological guide. This caveat should be stated more prominently in the text, since readers may otherwise over-read the cross-panel color comparison.","section":"Fig. 10"},{"comment":"The low-frequency calibration of sgnax is described only in prose; a table or figure showing the recovered-to-injected SNR ratio for the injections at 12.7 Hz and nearby frequencies, through the full whiten-resample-correlate path, would make the calibration auditable and would directly address the main concern raised in this report.","section":"Sec. VH"}],"recommendation":"major_revision","confidential_remarks":"For the editor: this is a solid software and validation paper with a potentially important calibration finding, but the production snax bug claim is stated more strongly than the evidence currently supports. The omission of a direct injection through the production chain and the post-hoc selection of the comparison band are the two issues that separate this from a clean accept. I do not see these as unfixable; they require additional targeted validation rather than a change of scope. The paper is likely a good fit for an instrument/software-oriented journal such as SoftwareX or a detector-characterization-focused submission."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read this. It's the most credible reimplementation of snax/omicron I've seen, and it earns a real referee.\n\nWhat's actually new: the unified dataflow graph on sgn, the measured 2.8x per-channel cost drop at 32 channels, the streaming glitch-rate frame channel, and the evidence that production snax over-reports SNRs below 25.6 Hz by sqrt(2)-2. The algorithms are inherited, but the consolidation and the operational products are new. The injection recovery (99.4%) is clean, and the omicron comparison on 24h of L1 strain is transparent: they report the 10-100 Hz, SNR>8 band where agreement is best (rate ratio 0.96, KS 0.02), but they also say full-band agreement is only within 12-20% with a near-threshold deficit above 2 kHz that they don't explain. That honesty is worth something.\n\nThe soft spot is the snax amplitude-error claim. The 0.705 median SNR ratio in the 12.8-25.6 Hz band is an inference from a cross-pipeline comparison, not a localization in the production code. The static audit didn't find the excess, and the band-stepped signature would look the same if sgnax's own low-rate whitening or resampler had a sqrt(2) gain error. The paper's evidence for the sgnax side is decent - injections down to 12.7 Hz through the full path recover a few percent - but it's reported in text without a figure, and the attribution to production snax is load-bearing for the headline claim. A direct injection through production snax would settle it. That said, the claim is plausible and the correction (divide by 1/sqrt(2) below 25.6 Hz) is safe even if the attribution wobbles, because it only makes the two pipelines consistent.\n\nMinor: the template-bank configuration for the injections is actually specified (Table II), so don't let that gripe stand. And the multi-channel scaling is on synthetic 4096 Hz channels on an M1 Pro - fine for a demonstration, not a production benchmark.\n\nBottom line: this is for the LVK detector-characterization community, and the pipeline deserves serious referee time. I'd send it to review with a request that the snax scaling claim either be verified by direct injection through production snax or softened to 'consistent with' rather than 'reveals.'","headline":"A credible unified rewrite of snax/omicron with clean injections and an honest but not fully closed case that production snax over-amplifies low-frequency SNRs.","tokens_in":30269,"tokens_out":3476,"would_cite":true,"duration_ms":26780,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"SGNAX reproduces both standard glitch searches in one streaming pipeline and exposes a $\\sqrt{2}$ SNR error in production features.","keywords":["gravitational-wave detector characterization","glitch triggers","matched filter","Q-transform","sine-Gaussian templates","streaming dataflow pipeline","SNR calibration","auxiliary channels"],"falsifier":"Feed a sine-Gaussian of analytically known optimal SNR at about 20 Hz through the production snax chain and through SGNAX; if the production chain reports roughly the analytic SNR while SGNAX reports a factor of $\\sqrt{2}$ lower, the scale error is confirmed, whereas if the two pipelines disagree differently across octave bands, the residual is in the comparison rather than in the production gain.","tokens_in":29205,"feed_emoji":"📡","tokens_out":10685,"duration_ms":80707,"temperature":0.7,"pith_summary":"SGNAX is presented as a single Python-native streaming pipeline that carries both standard detector-characterization analyses: the multi-rate sine-Gaussian matched-filter search of snax and the multi-resolution Q-transform excess-power search of omicron. On 24 hours of archival detector strain, the Q-transform mode reproduces the omicron trigger population in the 10–100 Hz band with a trigger-rate ratio of 0.96, a Kolmogorov–Smirnov statistic of 0.02, and a median per-trigger SNR difference of −2.4%, while the dense feature-stream mode recovers the snax loud-feature population with 80% per-bin coincidence. The same injection campaign that calibrates SGNAX shows that the production snax deployment inflates reported SNRs below 25.6 Hz by factors of $\\sqrt{2}$ to $2$, an error the paper localizes to a multiband amplitude gain rather than to the analysis statistic. If the claim is right, detector-characterization groups can run both analyses from one configuration and output schema, obtain Q-transform triggers at tens-of-seconds latency, and correct previously produced low-frequency feature SNRs.","feed_headline":"One pipeline reproduces both glitch searches and exposes a √2 SNR error","feed_subtitle":"Matches the standard trigger population to a few percent; calibration finds low-frequency SNRs inflated by √2 to 2.","key_machinery":"The load-bearing object is a single streaming dataflow graph whose per-channel branches share one data source and whitening stage, then diverge into either a multi-rate sine-Gaussian matched filter (quadrature correlation against a mismatch-spaced template bank, a per-window peak finder, and a cross-rate aggregator) or a Q-transform excess-power scan (bisquare-windowed tiles, a per-tile SNR estimator, cross-plane deduplication, and optional clustering). The shared front end is what makes per-channel cost fall as channels are added and what lets both algorithms see identical data and configuration in head-to-head comparisons. The second machinery element is the injection-calibrated scale reference: sine-Gaussian injections with analytically known optimal SNR pass through the full whiten-resample-correlate chain, establishing that SGNAX obeys the textbook matched-filter normalization and turning the residual against production snax into evidence of an amplitude error in that deployment rather than a difference in statistics.","core_discovery":"On its own terms, the paper establishes that a single streaming dataflow graph with a shared read-and-whitening front end can carry both the matched-filter and excess-power analyses without changing their statistics. The Q-transform branch reproduces the omicron trigger population on archival strain in the band most relevant to detector characterization—10–100 Hz with SNR above 8—at 96% of the reference trigger rate, with a 0.02 Kolmogorov–Smirnov statistic and a median per-trigger SNR 2.4% lower. The dense feature branch recovers the snax production loud-feature population with 80% per-bin coincidence and hourly rate correlation of 0.86. Injection calibration places the SGNAX SNR scale at the textbook matched-filter normalization, and the residual against production snax then localizes to a multiband amplitude error: rate bands below the production pipeline's 128 Hz resampling floor carry an extra gain of $\\sqrt{2}$ in 12.8–25.6 Hz and of $2$ below 12.8 Hz, so absolute production SNRs below 25.6 Hz should be divided by $\\sqrt{2}$ before comparison.","pith_inferences":["The step-like $\\sqrt{2}$ gain pattern at octave boundaries is a generic risk for any multi-rate whitened pipeline that resamples bands down to a common floor; an audit looking for SNR-ratio steps at resampling boundaries in other deployed pipelines would be a direct testable extension.","Because the scale error rescales SNRs without reordering them within a band, rank-based and threshold-based consumers are largely unaffected, but any absolute-SNR comparison crossing 25.6 Hz should show a systematic step; this predicts a discontinuity in production snax SNR histograms at exactly that frequency.","The high-frequency near-threshold residual above 2 kHz that the paper leaves open could be probed by running both modes with a common whitening filter; persistence of the residual would point to tiling or clustering rather than normalization.","Hosting both algorithms in one package turns the open problem of total-SNR recovery for broadband bursts into a per-pipeline configuration choice, which may accelerate a dedicated broadband estimator beyond the cluster-SNR improvement shown here."],"forward_implications":["Absolute SNRs recorded by production snax below 25.6 Hz carry the multiband amplitude excess and should be divided by $\\sqrt{2}$ (and by $2$ below 12.8 Hz) before being compared or combined with SGNAX features.","The Q-transform mode can deliver triggers within tens of seconds from shared-memory or streaming data sources, closing the gap between the batch cadence of production omicron deployments and the needs of detector commissioning.","A single configuration object and hdf5 output schema now serve both the matched-filter and Q-transform analyses, so downstream detector-characterization tools can consume both from one layout instead of two dependency stacks.","Processing many channels as parallel branches of one graph lowers per-channel cost by about 2.8× for the Q-transform between one and 32 channels, and a full day of 16 kHz strain is analyzed in about 14 minutes.","For band-limited white-noise bursts, a configurable non-overlapping cluster-SNR estimator recovers a median 0.66 of the injected SNR versus 0.48 for the single loudest tile, without changing localized sine-Gaussian recovery."],"supporting_citations":[{"why":"It defines the omicron Q-transform tiling, whitening normalization, and excess-power SNR estimator that the SGNAX Q-transform mode reimplements and against which the archival-strain comparison is made.","marker":"[13]"},{"why":"It defines the multi-rate sine-Gaussian matched-filter search whose statistic and dense feature mode SGNAX reimplements.","marker":"[18]"},{"why":"It documents the production feature-timeseries data product and low-latency role of snax that the SGNAX feature mode is compared with.","marker":"[19]"},{"why":"It supplies the streaming dataflow framework on which the unified single-graph architecture is built.","marker":"[22]"},{"why":"It provides the tensor-computation backend that evaluates the matched-filter correlations on CPU or GPU.","marker":"[23]"},{"why":"It establishes the event-trigger-generator validation methodology, the broadband-burst SNR under-recovery benchmark, and the comparison values that this paper extends.","marker":"[26]"},{"why":"It gives the sine-Gaussian metric from which the mismatch-driven template-bank spacing of both pipelines is derived.","marker":"[27]"},{"why":"It supplies the excess-power statistic that underlies the Q-transform SNR estimator.","marker":"[40]"}],"fun_headline_variants":["Unified glitch pipeline reproduces triggers and exposes √2 SNR error","One pipeline for matched-filter and Q-transform; finds √2 SNR inflation","SGNAX unifies glitch searches, calibrates SNRs to fix √2–2 error","Single dataflow graph reproduces omicron and snax, reveals SNR bias","Streaming pipeline matches standard trigger populations, pinpoints SNR error"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The diagnosis that production snax over-amplifies low-frequency SNRs assumes that SGNAX's own whitening and normalization chain is exactly the textbook matched-filter scale, so the measured 0.705 SNR ratio in the 12.8–25.6 Hz band is entirely the other pipeline's error rather than a difference in whitening, resampling, or template-bank construction.","fun_headline_variants_meta":{"raw":{"variants":["Unified glitch pipeline reproduces triggers and exposes √2 SNR error","One pipeline for matched-filter and Q-transform; finds √2 SNR inflation","SGNAX unifies glitch searches, calibrates SNRs to fix √2–2 error","Single dataflow graph reproduces omicron and snax, reveals SNR bias","Streaming pipeline matches standard trigger populations, pinpoints SNR error"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000291,"raw_usage":{"total_tokens":1798,"prompt_tokens":1139,"completion_tokens":659,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":755,"completion_tokens_details":{"reasoning_tokens":557}},"tokens_in":755,"tokens_out":659,"duration_ms":5533,"temperature":1.0,"reasoning_tokens":557,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T10:28:45.407914+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Feed a sine-Gaussian of analytically known optimal SNR at about 20 Hz through the production snax chain and through SGNAX; if the production chain reports roughly the analytic SNR while SGNAX reports a factor of $\\sqrt{2}$ lower, the scale error is confirmed, whereas if the two pipelines disagree differently across octave bands, the residual is in the comparison rather than in the production gain.","supporting_citations":[{"cited_title":"Robinet, Omicron: an algorithm to detect and char- acterize transient events in gravitational-wave detectors, Virgo Technical Document VIR-0545C-14 (2018)","cited_arxiv_id":null,"evidence_quote":"It defines the multi-rate sine-Gaussian matched-filter search whose statistic and dense feature mode SGNAX reimplements."},{"cited_title":"Robinetet al., Omicron: a Q-transform analysis tool for gravitational-wave detector data,https://git.ligo","cited_arxiv_id":null,"evidence_quote":"It documents the production feature-timeseries data product and low-latency role of snax that the SGNAX feature mode is compared with."},{"cited_title":"Godwin,Low-latency Statistical Data Quality in the Era of Multi-Messenger Astronomy, Ph.D","cited_arxiv_id":null,"evidence_quote":"It supplies the streaming dataflow framework on which the unified single-graph architecture is built."},{"cited_title":"Paszkeet al., PyTorch: An Imperative Style, High- Performance Deep Learning Library, inAdvances in Neu- ral Information Processing Systems 32(2019) pp","cited_arxiv_id":null,"evidence_quote":"It establishes the event-trigger-generator validation methodology, the broadband-burst SNR under-recovery benchmark, and the comparison values that this paper extends."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"It gives the sine-Gaussian metric from which the mismatch-driven template-bank spacing of both pipelines is derived."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"It supplies the excess-power statistic that underlies the Q-transform SNR estimator."}],"review_version":1}