{"id":"29da6fe0-c8c0-432f-83b2-73b655d8bcef","arxiv_id":"2605.23739","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"A wavelet-guided neural pipeline recovers previously known narrowband radio events from FAST observations of 33 exoplanet systems and reduces 139,127 detections to 803 veto-ready candidates; its one new candidate, toward K2-155, is argued to be terrestrial RFI.","lead":"This paper describes a new machine-learning pipeline that uses wavelet filtering and a lightweight neural network to find narrow drifting radio signals — possible signs of alien technology — in FAST telescope data from 33 exoplanet systems. It finds two candidate signals, one previously known and one new, and shows the new one is most likely human-made interference, not aliens.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The adopted τglob=1000 gate conflicts with reported recoveries: Fig. 9 shows detected signals with Global SNR 269–774 and Table 2 lists Kepler-438 XX at 659.6, all below the gate, implying the stated configuration would reject events the paper claims to recover.","rationale":"The reader's weakest_assumption focuses on calibration transfer from a single clean 1278–1380 MHz segment to full-band data. That is a legitimate generalizability risk, but it is an external robustness question: even if the network were trained on more diverse backgrounds, the central claim might still hold with recalibrated thresholds. The τglob=1000 inconsistency is more load-bearing because it is an internal contradiction: the pipeline as described cannot produce the very recovery results used to support the central claim. Table 2 and Fig. 9 are the primary evidence that the pipeline works on real data and on representative interference mixtures, and both contain entries with SNRglob below the stated gate. Whether this is a typo, a different gate application, or a real failure, it must be resolved before the pipeline's claimed performance can be accepted. The reader did mention this discrepancy in the rationale (issue 1), so we partially agree; however, I elevate it above the calibration-transfer concern because it directly questions whether the reported recoveries occurred under the specified configuration. The appropriate verdict remains CONDITIONAL: the paper needs to correct or clarify this inconsistency, but the overall approach is not fatally undermined. Therefore no change to the reader's verdict is needed.","tokens_in":24091,"tokens_out":5036,"duration_ms":57982,"concrete_test":"Use the released WaveSETI code to run the exact stated configuration (τglob=1000, θconf=0.7, θIoU=0.9) on the Kepler-438 (NBS 210629) and HD 180617 (NBS 210421) observations for XX-only, YY-only, and XX+YY inputs. Record which Table 2 detections survive each stage. Additionally, for the Fig. 9 simulated cases, compute the patch-level SNRglob and check whether any detection would survive the gate. If the low-SNRglob rows or Fig. 9 examples are rejected, the paper's operating point must be revised or the claim re-scoped; if they survive, the text must clarify the actual gate definition (per-patch vs per-detection, or a different τglob).","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central real-data claim is that the pipeline recovers the Kepler-438 and HD 180617 events and yields 803 veto-ready candidates under a fixed auditable operating point. But the operating point is internally inconsistent: §2.3.2 defines a patch-level Global SNR gate (SNRglob < τglob rejected) and §3.1 states τglob = 1000. Table 2 lists this-work detections with SNRglob values of 659.6 (Kepler-438 XX), 1175.6 (HD 180617 XX), and 1116.6 (Kepler-438 YY). If the gate is applied as described, the XX detections with SNRglob < 1000 cannot survive, yet they are presented as recovered. Similarly, Fig. 9's robustness gallery shows successful detections on cleaned maps labeled Global SNR 268.88, 370.67, and 774.46 — all below τglob = 1000 — so those examples would be rejected by the stated gate. This is not a question of transferability or noise statistics; it is a direct contradiction between the specified threshold and the evidence used to support the pipeline's sensitivity. If the gate is actually applied, the two anchor recoveries are not reproduced under the claimed configuration, weakening the strongest claim. If the gate is not applied as written (e.g., a typo such as 100 instead of 1000, or a per-detection rather than per-patch definition), then the 'auditable threshold control' is not accurately reported, and the 803-candidate yield cannot be traced to the stated parameters. Either way, the central claim is not verifiable from the text as it stands.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents MSWNet, a wavelet-integrated encoder–decoder followed by a lightweight endpoint regressor, as a staged inference pipeline for narrowband technosignature searches in FAST L-band dynamic spectra. Training data are generated by injecting synthetic drifting signals (linear and quadratic, |ν̇|≤4 Hz/s, widths 1–3 channels) into a clean 1278–1380 MHz background segment. At inference, the pipeline applies a robust global-SNR patch gate, confidence gating with NMS, cross-patch stitching, raw-data S/N validation, and a 19-beam anticoincidence veto. On the 33-target FAST campaign, the authors report recovering two previously published events (Kepler-438/NBS 210629 and HD 180617/NBS 210421) and producing 803 veto-ready candidates from 139,127 single-beam detections. A newly identified candidate, NBS 260108 toward K2-155, is analyzed in detail and attributed to likely RFI on the basis of polarization asymmetry, cross-target recurrence, and ensemble periodicity diagnostics.","tokens_in":24501,"tokens_out":8241,"duration_ms":97598,"significance":"If the reported results are reproducible, the paper offers a credible ML-based alternative to drift-grid searches, with attractive properties: an interpretable wavelet front end, explicit staged thresholds, open-source code, and real-data anchoring against two prior peer-reviewed FAST detections. The use of prior published events as recovery tests is a legitimate validation strategy, and the compact 803-candidate output directly addresses the human-review bottleneck that dominates current SETI pipelines. However, the paper's central verification claim is currently blocked by an internal inconsistency between the stated operating point and the reported recovery values. The significance of the contribution is therefore conditional on resolving that inconsistency.","major_comments":[{"comment":"The stated operating point is internally inconsistent. Section 2.3.2 defines a patch-level rejection rule SNRglob < τglob, and §3.1 states τglob = 1000. Under this rule, Table 2's Kepler-438 XX recovery (SNRglob = 659.6) cannot survive the first post-processing stage, and Fig. 9 shows successful detections labeled Global SNR = 268.88, 370.67, and 774.46, all below τglob = 1000. Either the gate is not actually 1000, the recoveries are produced with a different configuration, or the SNRglob values in Table 2/Fig. 9 are not the gate statistic. The §4.1.4 caveat that events are recovered 'before downstream screening' does not resolve this, because the global-SNR gate is upstream of the later screening stages in the stated pipeline. Please clarify the exact configuration used for each reported result and provide per-stage counts (patches passing the gate, detections, events, candidates) under","section":"§2.3.2, §3.1, Table 2, Fig. 9"},{"comment":"The pipeline's fixed thresholds are calibrated on simulations built from a single clean 1278–1380 MHz segment, but are then applied across the full 1.05–1.45 GHz band. The claim that τglob transfers because SNRglob is MAD-normalized is plausible but not demonstrated. Figure 4 itself shows strong frequency-dependent structure in the real data, including clustering at known interference sub-bands. As written, the 803-candidate yield cannot be separated from the choice of τglob, and the reader cannot tell whether the gate is over- or under-rejecting outside the training band. Please provide a quantitative calibration check, e.g., injection-recovery rates in subbands outside 1278–1380 MHz, or a comparison of SNRglob distributions and noise-floor stability across the full band.","section":"§2.1, §3.1, Fig. 4"},{"comment":"The real-data evaluation rests on two anchored recoveries and a final candidate count, but no end-to-end completeness or contamination measurement is reported. The two anchors demonstrate sensitivity to previously detected events, but they do not quantify recall (fraction of injected signals recovered at the stated thresholds) or precision (false-positive rate before and after veto) on real FAST data. To support the claim that this is a practical replacement for drift-grid search, please include injection-recovery tests on real observations spanning the full band, and report the number of events rejected at each veto stage under the nominal configuration.","section":"§3.1, §4.1.4"}],"minor_comments":[{"comment":"The figure caption and axis labels contain garbled glyph sequences in the provided version (e.g., '/uni00000013/...'), making the threshold-sensitivity figure unreadable. Please regenerate the figure and caption.","section":"Figure 3"},{"comment":"The raw-data S/N cutoff is given as 'e.g., S/N>10'. Please state the exact value used for the 803-candidate result, since the recovery rows in Table 2 include S/N values below 10 (e.g., 4.4 and 7.8) and the text later clarifies these are before downstream screening.","section":"§2.3.2"},{"comment":"Typo: 'Npredections' should be 'N_predictions'. Also, Eq. (8) should define n and k precisely, since k depends on n but the patch area is not defined.","section":"§2.2"},{"comment":"The phrase 'recovered by our pipeline before downstream screening' is important and should appear in the abstract or Section 3 summary to avoid the impression that the two anchor events pass all veto stages.","section":"§4.1.4"},{"comment":"The code link (WaveSETI9) should include a version/commit and a statement of the exact inference configuration used for the reported numbers, to support reproducibility.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The overlap in authorship with the prior Tao et al. (2022) and Luan et al. (2023) papers is worth noting, but the anchored events are peer-reviewed, published detections on real telescope data, so using them as recovery tests is a legitimate validation choice. The main blocker is the threshold inconsistency: as written, the claimed operating point is not auditable. This is fixable by re-running or clearly reporting the actual gate configuration and per-stage counts. The paper does not need a fundamentally different evaluation, but it does need a consistent, reproducible statement of its own thresholds."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this paper is a worthwhile contribution to SETI pipeline work, but it has an internal inconsistency about its reported detection threshold that needs to be sorted out before the main claims are credible. The authors describe a staged pipeline—wavelet U-Net with cached detail coefficients, a lightweight endpoint regressor, confidence gating, and a 19-beam veto—and apply it to FAST data. The two anchored recovery tests (Kepler-438 and HD 180617) match published frequencies and drift rates closely, and the funnel from 139,127 detections to 803 candidates is transparent. The treatment of the new K2-155 event is a model of honest RFI investigation: single-polarization, same-session recurrence, comb structure, and explicit admission that the coupling mechanism remains unknown. Code is released. That is real work, and the community can use it.\n\nThe soft spots are not fatal, but they are real. The most pressing is the τglob threshold. Section 3.1 states τglob = 1000, but Table 2 lists the recovered Kepler-438 XX detection with SNRglob = 659.6. If the gate is applied as written, that event would be rejected before it can become a candidate. Figure 9 also shows successful detections with global SNR values of 268–774, all below the stated gate. Either the reported threshold is a typo, the SNRglob is computed differently than described, or the recoveries come from a configuration that does not match the stated operating point. The paper needs to fix this, because the 'auditable threshold control' claim is central.\n\nBeyond that: there is no quantitative comparison to TurboSETI on the same data, so 'compact candidate set' is not yet demonstrated relative to a baseline. The calibration on a single clean 1278–1380 MHz segment raises transferability concerns across the full 1.05–1.45 GHz band. The ensemble diagnostics (R values, 12.3 kHz spacing) lack significance estimates. Each of these is addressable, and none undercuts the basic architecture.\n\nWho should read it: anyone building ML-based SETI pipelines, especially for FAST or other multibeam instruments. It deserves a serious referee, with a request to resolve the threshold inconsistency, add a baseline comparison, and quantify the significance of the ensemble statistics. If those are fixed, I would cite it and likely use it as a reference pipeline.","headline":"Useful SETI pipeline paper, but the reported τglob=1000 gate contradicts the SNRglob values in its own Table 2, so the operating point is not yet trustworthy.","tokens_in":25196,"tokens_out":3596,"would_cite":true,"duration_ms":39564,"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 wavelet-based ML pipeline recovers known FAST narrowband events and reduces 139,127 detections to 803 inspectable candidates.","keywords":["technosignatures","SETI","wavelet analysis","neural networks","narrowband signal search","FAST telescope","drift-rate search","radio frequency interference"],"falsifier":"On a fresh FAST observation with a known injected narrowband signal at S/N≈10 and drift within ±4 Hz/s, run the pipeline with thresholds frozen; if the signal is missed while a conventional drift-grid search at the same S/N finds it, the claim of transferable sensitivity across RFI environments is falsified.","tokens_in":23820,"feed_emoji":"📡","tokens_out":9212,"duration_ms":94364,"temperature":0.7,"pith_summary":"The paper claims that narrowband technosignature searching can be reframed from an exhaustive drift-rate grid scan into a two-stage operation: a wavelet-based network (MSWNet) cleans the time–frequency image, and a lightweight estimator regresses each signal's start and end frequencies directly. The payoff, demonstrated on real FAST observations of 33 exoplanet systems, is a compact, auditable candidate list: 139,127 single-beam detections are reduced to 803 veto-ready candidates, and two previously reported events are recovered with consistent frequency and drift. The paper also identifies and then diagnoses a new drifting narrowband signal toward K2-155 (1148.4167 MHz, −0.038 Hz/s, S/N≈15), concluding on the evidence that it is most plausibly anthropogenic or instrumental rather than extraterrestrial. A sympathetic reader would care because the bottleneck in modern SETI is verification, not detection, and this pipeline moves the decision boundary to a staged, inspectable sequence.","feed_headline":"Pares 139,127 SETI detections to 803 candidates with wavelet search","feed_subtitle":"A wavelet net plus endpoint regression cuts the review burden, and the one new candidate is traced to interference.","key_machinery":"The central object is MSWNet (Multi-Scale Wavelet Net), an encoder–decoder in which every pooling operation is replaced by a two-dimensional discrete wavelet transform (DWT2D). At each scale the encoder caches the detail bands (LH/HL/HH) and the decoder reuses them through the inverse transform, so fine-scale structure is carried rather than hallucinated; this preserves weak narrowband tracks while suppressing broadband and impulsive interference. Downstream, a lightweight parameter estimator regresses (f_start, f_stop) using a 2×N_f representation, and a staged post-process — global-SNR gate, confidence gating plus non-maximum suppression, cross-patch stitching, raw-data S/N validation, and","core_discovery":"Central claim: narrowband technosignature search can be recast as wavelet-guided feature extraction plus endpoint regression instead of an exhaustive drift-rate scan. On FAST L-band observations of 33 exoplanet systems, MSWNet cleans each patch, a lightweight estimator regresses (f_start, f_stop) and confidence, and staged filters plus a 19-beam veto reduce 139,127 detections to 803 candidates. Two earlier events (Kepler-438, HD 180617) are recovered with consistent frequency and drift. A new drifting signal toward K2-155 (1148.4167 MHz, −0.038 Hz/s, S/N≈15) is diagnosed as likely RFI via polarization, cross-target recurrence, and 12.3 kHz spacing.","pith_inferences":["If the calibration premise holds, the strongest near-term gain is to swap the endpoint regressor for a physics-driven track fitter (e.g., a Hough-style line/curve detector) on MSWNet's cleaned maps; the paper itself flags estimator localization error as the main veto risk.","The survey-wide ~12.3 kHz periodicity in the matched-control population suggests an instrumental lattice that could be characterized in advance and subtracted, which would shrink the candidate list further.","A cheap falsification of the transferability claim would be to run the frozen thresholds on a different telescope band containing an injected S/N≈10 drifting signal; success or failure would calibrate how much of the pipeline depends on this specific instrument's noise bed.","The wavelet-path ablation suggests MSWNet's cleaned maps could serve as a general RFI-mitigation front end for other narrowband radio analyses, not just SETI searches."],"forward_implications":["Because detection no longer enumerates a drift-rate grid, output volume is decoupled from the size of the drift hypothesis set; retained yield is controlled by explicit thresholds that can be audited and adjusted.","The same pipeline, with fixed thresholds, recovers the Kepler-438 and HD 180617 events found by earlier analyses of the same FAST data, showing sensitivity to the same class of narrow drifting features.","Multi-beam anticoincidence veto plus polarization and recurrence diagnostics remain essential: the new K2-155 candidate survived drift- and beam-based filters but was classified as likely RFI on the broader evidence chain.","The 803 veto-ready candidates form a compact input for human or automated expert inspection, directly addressing the verification bottleneck.","Ablation results imply that removing the wavelet high-frequency pathway would push marginal S/N≈10 signals below detectability, so the detail path is part of the sensitivity budget."],"fun_headline_variants":["Wavelet pipeline cuts SETI detections from 139k to 803 on FAST data","Wavelet search on 33 exoplanets yields 803 veto-ready candidates","Wavelet regression narrows 139k detections to 803 candidates","FAST wavelet search trims detections from 139k to 803","Wavelet pipeline recovers known events, flags one RFI in 33 stars"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that one clean strip of the observed band plus simulated signals captures the noise and radio-frequency-interference behavior of the full 1.05–1.45 GHz observation; if interference elsewhere looks different, or real signals fall outside the simulated drift/curvature/width range, the pipeline's thresholds are miscalibrated and the 803-candidate yield is not trustworthy.","fun_headline_variants_meta":{"raw":{"variants":["Wavelet pipeline cuts SETI detections from 139k to 803 on FAST data","Wavelet search on 33 exoplanets yields 803 veto-ready candidates","Wavelet regression narrows 139k detections to 803 candidates","FAST wavelet search trims detections from 139k to 803","Wavelet pipeline recovers known events, flags one RFI in 33 stars"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000554,"raw_usage":{"total_tokens":2462,"prompt_tokens":717,"completion_tokens":1745,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":461,"completion_tokens_details":{"reasoning_tokens":1641}},"tokens_in":461,"tokens_out":1745,"duration_ms":13314,"temperature":1.0,"reasoning_tokens":1641,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T13:19:28.507527+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a fresh FAST observation with a known injected narrowband signal at S/N≈10 and drift within ±4 Hz/s, run the pipeline with thresholds frozen; if the signal is missed while a conventional drift-grid search at the same S/N finds it, the claim of transferable sensitivity across RFI environments is falsified.","supporting_citations":[],"review_version":3}