{"id":"4f882077-e17a-4bb1-a1d5-d8248b8833da","arxiv_id":"2507.04339","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A template-matching radio trigger keeps 90% of simulated air-shower signals while rejecting over 75% of background pulses, and an array-level trigger adds direction and polarization consistency checks.","lead":"This paper develops a new radio trigger for the GRAND neutrino observatory that recognizes air-shower signals by pulse shape at the antenna and checks timing, direction, and polarization across the array. It reports that the shape-based trigger rejects more than three quarters of background pulses while keeping 90 percent of simulated signals, an important step toward running a 200,000-antenna array.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 75% background rejection at 90% signal efficiency is measured on simulated pulses from the same library used to select the templates; it may not transfer to real air showers.","rationale":"The paper's headline number is an offline efficiency measured on simulated signals. The same simulation library supplies both the five templates and the signal sample used to set the efficiency point, so the comparison of signal and background rho distributions is, at its core, a comparison between a simulation-tuned matched filter and real RFI pulses. The only real-signal test is the brief case study in Section 3.2, which is qualitative and does not use the FIR-filtered pipeline or a threshold derived from the background database. This is the weakest link in the chain from the reported 75%/90% performance to the designed rate reduction. I agree with the reader's weakest_assumption; the baseline comparison ambiguity (Section 3.2, stars in Fig. 1) is secondary, because even a fully correct baseline comparison would not rescue the claim if real signals do not match the templates. The proposed tests are internal and external: a held-out simulation split bounds the template-selection bias, and a real-event rho retention measurement at the background-derived threshold directly tests transferability. The conditionality of the reader's verdict is appropriate: the method is standard, the background rejection is measured on real data, and the online throughput has a 5x margin, but the signal-efficiency claim requires external validation before the numbers can be relied on for the design.","tokens_in":10014,"tokens_out":14194,"duration_ms":136514,"concrete_test":"Select a set of independently identified real air-shower events in GP300 archive data, for example the cosmic-ray candidate CD events of [11], or events with at least 5 DUs passing a plane-wave fit and polarization consistency in a CD run. Process them with the exact offline FLT-1 pipeline (FIR filter, FLT-0-light, the five templates of Section 3.1) and record the best-match rho. Set rho_thresh at the value that gives at least 75% background rejection on the FLT background MD database, then measure the signal retention of the real candidates; if fewer than 90% pass, the central claim does not transfer. Additionally, re-run the Section 3.2 efficiency on a disjoint half of the simulation library, with templates selected on the other half, to quantify in-sample optimism.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central FLT-1 claim (Section 3.2, Conclusion) is that, for SNR > 5, a 90% signal-selection efficiency is accompanied by >=75% background rejection. Both quantities come from databases built in Section 2.2: the five ADC templates are selected from simulated ZHAireS pulses propagated through GRANDlib (Section 3.1), and the signal database used to set the rho_thresh threshold and to measure the 90% selection efficiency is drawn from the same 25,000-event simulation library. The background rejection is measured on real GP300 MD pulses, so the ROC separation rests entirely on simulated signal shapes matching real air-shower ADC pulses. The paper's real-data checks do not close this gap: the transformer versus cosmic-ray candidate case study (Section 3.2) is qualitative, is run without the FIR filter used in the nominal pipeline, and does not compare real candidates against the background-derived threshold. The 225 candidate pulses are not independently confirmed air showers, and the good agreement between the simulated RF chain and data (Section 2.2) validates the electronics, not the ZHAireS electric-field waveforms. If real pulses differ in shape, for example for near-horizontal geometries or due to unmodeled direction-dependent antenna response, the threshold that retains 90% of simulated signals may retain fewer real signals, or the corresponding background rejection may be below 75%, invalidating the headline rate reduction from 1 kHz to 100 Hz.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper presents the NUTRIG trigger development for the GRAND radio detector. A first-level trigger (FLT-1) uses five ADC pulse templates, derived from ZHAireS air-shower simulations propagated through the GRANDlib RF chain, and computes the maximum cross-correlation ρ between an input pulse and the templates; a threshold on ρ is intended to reject radio-frequency background while retaining air-shower signals. The authors report that for SNR > 5 a 90% signal-selection efficiency on simulated pulses is accompanied by ≳75% background rejection on real GP300 monitoring data, compared with ≲20% rejection for the current FLT-0-optimal trigger, and that the algorithm runs online at ~5.5 kHz on the DU CPU. A second-level trigger (SLT) applies cuts on timing, direction, and polarization, with individual background rejections of 82%, 98%, and 23% at 95% signal-selection efficiency for events with ≥5 DUs.","tokens_in":10311,"tokens_out":14499,"duration_ms":140715,"significance":"If the claimed performance transfers from simulations to sky data, FLT-1 would reduce the DU-level trigger rate from 1 kHz to 100 Hz as designed, a critical step for GRAND's planned 10^4–10^5 DU arrays. The paper's strengths include the use of real GP300 MD and CD data for background rejection estimates, the explicit online throughput measurement on the actual DU CPU, and the transparent definition of the trigger logic and cuts. The SLT azimuth reconstruction result is a promising basis for the array-level trigger. The main risk is that the signal-side efficiencies are measured on simulated pulses from the same library used to build the templates, and no disjoint validation set or real air-shower pulse sample is used to confirm the 90% operating point.","major_comments":[{"comment":"The five ADC templates are selected from the same 25,000-event ZHAireS/GRANDlib simulation library that is used to build the FLT signal database, and the ρ_thresh threshold is tuned on that database to achieve a 90% signal-selection efficiency. No train/test split or cross-validation is described, so the 90% efficiency is a training-set measurement and the accompanying ≥75% background rejection is not an unbiased estimate of online performance. Please evaluate the template selection and threshold on a disjoint set of simulated pulses (or use k-fold cross-validation) and report the resulting efficiencies and their statistical uncertainty.","section":"§3.1 and §3.2"},{"comment":"The comparison to FLT-0-optimal is not specified clearly. The text says FLT-0-optimal is applied directly to the MD background database and that selection and rejection efficiencies are the fractions of accepted and rejected traces; since the MD traces are background, this measures background acceptance, not signal-selection efficiency. The subsequent statement that the comparison is made \"for the same signal selection efficiency achieved by the FLT-0-optimal\" requires a measurement of that signal efficiency that is not described. Please state explicitly how the FLT-0-optimal signal-selection efficiency was obtained (e.g., from the simulated signal database) and give the operating point used for the comparison.","section":"§3.2"},{"comment":"The FLT signal database is constructed by first applying the FLT-0-light trigger, so the reported 90% signal-selection efficiency of FLT-1 is conditional on passing FLT-0-light. The total signal-selection efficiency of the FLT-0-light+FLT-1 chain is the product of the FLT-0-light efficiency and the FLT-1 conditional efficiency, but the former is not quoted. Please report the end-to-end signal efficiency and the corresponding end-to-end background rate, or explicitly justify that FLT-0-light passes all air-shower pulses in the considered SNR range.","section":"§2.2 and §3.2"},{"comment":"The real-data case study with the 225 cosmic-ray candidate pulses does not close the simulation-to-data gap: the candidates are not independently confirmed air showers, no ρ_thresh is applied to them, and the comparison to the background-derived threshold is not shown. Please report the distribution of ρ for the candidate pulses relative to the chosen ρ_thresh and, if possible, validate the signal-selection efficiency on an independent sample of identified air-shower pulses (e.g., from a different simulation code or from events reconstructed by the SLT).","section":"§3.2 and §2.1"}],"minor_comments":[{"comment":"The sentence \"In this case study, no FIR filter is applied to choose the 5 optimal templates (Section 3.1)\" is ambiguous and appears to contradict Section 3.1, where the templates are selected after applying the FIR filter. Please clarify whether the case study applies the same FIR-filtered templates to unfiltered data or uses a separate set of unfiltered templates.","section":"§3.2"},{"comment":"The right panel is described as \"Signal selection efficiencies versus background rejection efficiencies as a function of SNR\"; please clarify whether each curve is an ROC-like curve obtained by varying ρ_thresh or a set of single operating points per SNR bin.","section":"Fig. 1 and §3.2"},{"comment":"Please state whether the Δθ and Δφ cuts are applied independently or jointly, and report the combined signal-selection and background-rejection efficiencies if they are applied together.","section":"§4.2"},{"comment":"The column header \"Number of SNR\" is unclear; it should read \"SNR range\" or \"SNR bin\".","section":"Table 1"},{"comment":"Please specify the conditions of the throughput measurement, in particular whether the quoted ~5.5 kHz includes the FIR filtering, FLT-0 processing, and cross-correlation over all five templates and two polarizations, and whether the CPU was otherwise idle.","section":"§3.3"},{"comment":"The statement \"e.g., ∼90% compared to ≲20% for 7≤SNR<8\" compares the FLT-1 background rejection at 90% signal selection with the FLT-0-optimal background rejection at its own (possibly different) signal-selection efficiency; add a sentence stating the signal-selection efficiency at which the FLT-0-optimal value is evaluated.","section":"§3.2"}],"recommendation":"major_revision","confidential_remarks":"This is a proceedings paper, and the central idea is promising and within the scope of the venue. The main issue is that the headline quantitative claim is measured on the same simulation library used to derive the templates and threshold; a revision that adds a validation split and clarifies the FLT-0-optimal comparison would make the claim sound. I do not see grounds for rejection, but the current text as written overstates the readiness of the 90%/75% numbers."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Hi X,\n\nQuick take: this is a genuine step forward for GRAND's trigger development, with the first quantitative FLT-1 results on GP300 data and a real online throughput test on the FEB CPU. The headline number—≥75% background rejection at 90% signal selection for SNR>5—is new, but it rests on simulated pulses drawn from the same ZHAireS library that provided the five templates. That is a self-consistency check, not yet a closed proof for real air showers.\n\nWhere it earns credit: the background side is real data (6,628 MD pulses, 2,146 CD events), the online test at ~5.5 kHz is concrete engineering, and the SLT section is honest about the polarization step's weakness (23% rejection) and the unfinished status. The transformer-vs-cosmic-ray-candidate case study is suggestive, and the authors don't oversell it.\n\nSoft spots, in order of importance. First, no train/test split is described: the templates and the signal-efficiency test both come from the same 25,000-event simulation library. The 90% selection efficiency therefore measures how well the templates match the library, not necessarily the sky. The RF-chain validation against data covers the electronics but not the ZHAireS electric-field waveforms; near-horizontal geometries or direction-dependent antenna response could change ADC pulse shapes. This should be stated as a caveat and addressed with real cosmic-ray showers or a split library before the rate-reduction claim is treated as baseline. Second, the FLT-0-optimal comparison in Section 3.2 is confusing: the text says it was applied to the MD background data, then references signal selection efficiency—probably both databases were used, but it needs rewording. Third, the case study runs with no FIR filter and no confirmation that the 225 candidates are actual air showers; fine as an illustration, but qualitative. Also there are no uncertainties reported anywhere; for a proceedings that is understandable, but for the headline efficiency it would help.\n\nOverall this is a credible status report, not a finished performance claim. It belongs in ICRC proceedings and deserves a serious referee for a journal version if the simulation-to-data gap is addressed. I'd send it to review.","headline":"Solid engineering status report for GRAND's NUTRIG trigger with new quantitative results, but the headline rejection efficiency is measured on simulated signals from the same library used to build the templates, so treat the 1 kHz to 100 Hz claim as conditional on simulation-to-data agreement.","tokens_in":10911,"tokens_out":2614,"would_cite":true,"duration_ms":28936,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A template-matching pulse-shape trigger rejects more than 75% of background radio noise while keeping 90% of air-shower signals at SNR > 5.","keywords":["radio self-trigger","air-shower detection","template matching","first-level trigger","GRAND","background rejection","cross-correlation","neutrino detection"],"falsifier":"Deploy the FLT-1 at GP300, feed it the real cosmic-ray candidate pulses and forced-trigger background pulses from the same run, and compute the background rejection at the fixed 90% signal-selection threshold; if it does not reach $\\gtrsim 75\\%$ for SNR $> 5$ on sky data, the central claim collapses.","tokens_in":9790,"feed_emoji":"📡","tokens_out":4108,"duration_ms":41889,"temperature":0.7,"pith_summary":"The paper argues that a first-level radio trigger which matches the shape of an ADC pulse against a small library of air-shower templates can cleanly separate real cosmic-ray signals from radio-frequency interference at the detection-unit level. It reports a background rejection efficiency of at least 75% while holding the signal-selection efficiency at 90% for pulses with signal-to-noise ratio above 5, compared with the current GP300 double-threshold trigger, which rejects under 20% of background at higher SNR. If this holds on sky data, the per-detector trigger rate can drop from 1 kHz to the 100 Hz design target, and a second-level array trigger built on timing, direction, and polarization reconstructions can reduce the array trigger rate from 10 Hz to 1 Hz.","feed_headline":"New radio trigger rejects 75% of background, keeps 90% of signals","feed_subtitle":"Template-matched pulse shapes could cut GRAND's per-detector trigger rate from 1 kHz to 100 Hz.","key_machinery":"The central object is the best-match cross-correlation value $\\rho = \\max_{i,j,\\tau} |\\rho_{ij}(\\tau)|$, where $\\rho_{ij}(\\tau) = \\int T_{ij}(t) V_i(t+\\tau)\\, dt$ is the cross-correlation between an input trace $V_i$ and template $T_{ij}$ over a sliding 200-ns window. The five templates are chosen as the most representative simulated ADC pulses in five bins of the opening angle to the shower axis, after propagation through the antenna and RF chain, and the absolute value handles opposite-polarity pulses. This single scalar discriminates air-shower pulses from background.","core_discovery":"The paper claims that a template-matching first-level trigger, called FLT-1, achieves a background rejection efficiency of $\\gtrsim 75\\%$ at a fixed signal-selection efficiency of 90% for pulses with SNR $> 5$, using only the maximum absolute cross-correlation between a 200-ns ADC trace and one of five air-shower templates. This is a substantial improvement over the current GP300 FLT-0 double-threshold trigger, which rejects $\\lesssim 20\\%$ of background for SNR $> 7$. The paper also reports that the FLT-1 algorithm, implemented on the detection-unit CPU, processes events at up to about 5.5 kHz, well above the expected input rate of 1 kHz, and that an array-level second-level trigger combining plane-wave-front timing, conic signal-strength footprint, and polarization cuts reaches 95% signal-selection efficiency while rejecting 82% of background in zenith angle, 98% in azimuth angle, and 23% in polarization.","pith_inferences":["Editorial: the 90% signal-selection efficiency is measured on simulated pulses drawn from the same simulation library that produced the templates, so the real test is how the templates transfer to genuine cosmic-ray pulses once deployed at GP300.","Editorial: the qualitative separation between electric-transformer pulses and cosmic-ray candidates suggests that high-frequency RFI can be rejected purely by pulse shape, making the method robust against the dominant local noise source at the site.","Editorial: the azimuth-angle reconstruction being the strongest SLT separator implies that a footprint-based direction cut could carry most of the array-level rejection even if the polarization step remains weak.","Editorial: a testable extension would be to record real $Z$-channel pulses and build a matching template bank, since the RF chain for that polarization differs from the $X$ and $Y$ channels."],"forward_implications":["If correct, the FLT-1 would allow the detection-unit trigger rate to be reduced from the nominal 1 kHz to the 100 Hz design target, a factor of ten drop in online data flow.","The measured processing speed of about 5.5 kHz leaves sufficient headroom over the expected 1 kHz input, so the trigger can run on the existing detection-unit CPU without additional hardware.","The SLT cuts on azimuth, zenith, and polarization, combined with FLT-1 rejection, could reduce the array-level central trigger rate from 10 Hz to 1 Hz as designed.","The method is directly extendable to the vertical $Z$ polarization by selecting equivalent templates, which would add a third independent polarization channel for rejection."],"supporting_citations":[{"why":"Supplies the GRANDlib simulation chain that propagates simulated electric fields through the antenna and RF chain to the ADC, generating the signal templates.","marker":"[10]"},{"why":"Provides the ZHAireS air-shower simulations from which the 25,000-event library and the five ADC templates are derived.","marker":"[12]"},{"why":"Inspires the cross-correlation template-matching approach used to compute the best-fit value $\\rho$.","marker":"[13]"},{"why":"Describes the LPNHE test bench where the online implementation of FLT-1 was ported and tested at rates up to about 5.5 kHz.","marker":"[14]"},{"why":"Introduces the conic method used in the second-level trigger to fit an elliptical signal-strength footprint and reconstruct zenith and azimuth angles.","marker":"[9]"},{"why":"Provides the analytic plane-wave-front reconstruction used to estimate arrival direction and reject timing outliers in the SLT.","marker":"[15]"},{"why":"Supplies the polarization-estimation method that the SLT adapts to voltage amplitudes for the polarization selection cut.","marker":"[16]"},{"why":"Identifies the electric transformer and airplanes as dominant RFI sources and provides the cosmic-ray candidate pulses used in the case study.","marker":"[11]"}],"fun_headline_variants":["Template matching radio trigger cuts noise 75%, keeps 90% of signals","Radio self-trigger for GRAND: 75% noise cut at 90% signal efficiency","CPU trigger for GRAND processes 5.5 kHz events, exceeds 1 kHz input","New radio self-trigger: 75% background rejection with 90% signal retention","Array-level radio trigger for GRAND: 82% zenith, 98% azimuth noise cut"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The template library is built from simulated air showers propagated through the simulated RF chain, and the reported rejection efficiency relies on those templates matching what real air-shower pulses look like at the ADC.","fun_headline_variants_meta":{"raw":{"variants":["Template matching radio trigger cuts noise 75%, keeps 90% of signals","Radio self-trigger for GRAND: 75% noise cut at 90% signal efficiency","CPU trigger for GRAND processes 5.5 kHz events, exceeds 1 kHz input","New radio self-trigger: 75% background rejection with 90% signal retention","Array-level radio trigger for GRAND: 82% zenith, 98% azimuth noise cut"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001391,"raw_usage":{"total_tokens":5607,"prompt_tokens":904,"completion_tokens":4703,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":520,"completion_tokens_details":{"reasoning_tokens":4587}},"tokens_in":520,"tokens_out":4703,"duration_ms":30916,"temperature":1.0,"reasoning_tokens":4587,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T19:50:11.244809+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Deploy the FLT-1 at GP300, feed it the real cosmic-ray candidate pulses and forced-trigger background pulses from the same run, and compute the background rejection at the fixed 90% signal-selection threshold; if it does not reach $\\gtrsim 75\\%$ for SNR $> 5$ on sky data, the central claim collapses.","supporting_citations":[{"cited_title":"Alves Batistaet al","cited_arxiv_id":null,"evidence_quote":"Supplies the GRANDlib simulation chain that propagates simulated electric fields through the antenna and RF chain to the ADC, generating the signal templates."},{"cited_title":"Álvarez-Muñiz, W","cited_arxiv_id":null,"evidence_quote":"Provides the ZHAireS air-shower simulations from which the 25,000-event library and the five ADC templates are derived."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Inspires the cross-correlation template-matching approach used to compute the best-fit value $\\rho$."},{"cited_title":"CorreaPoSICRC2023(2023) 990","cited_arxiv_id":null,"evidence_quote":"Describes the LPNHE test bench where the online implementation of FLT-1 was ported and tested at rates up to about 5.5 kHz."},{"cited_title":"Köhleret al","cited_arxiv_id":null,"evidence_quote":"Introduces the conic method used in the second-level trigger to fit an elliptical signal-strength footprint and reconstruct zenith and azimuth angles."},{"cited_title":"Ferrière et al","cited_arxiv_id":null,"evidence_quote":"Provides the analytic plane-wave-front reconstruction used to estimate arrival direction and reject timing outliers in the SLT."},{"cited_title":"Emergences","cited_arxiv_id":null,"evidence_quote":"Supplies the polarization-estimation method that the SLT adapts to voltage amplitudes for the polarization selection cut."},{"cited_title":"Lavoisieret al","cited_arxiv_id":null,"evidence_quote":"Identifies the electric transformer and airplanes as dominant RFI sources and provides the cosmic-ray candidate pulses used in the case study."}],"review_version":1}