{"id":"ea9bd39e-d1fa-4042-9a36-64efbe503826","arxiv_id":"2607.28404","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Virtual ANF frequency-track statistics plus FFT features raise family-level GNSS RFI classification accuracy on compact gradient-boosted trees across synthetic and recorded datasets.","lead":"The paper shows that cheap statistics from a virtual adaptive notch filter, combined with ordinary FFT features, improve real-time GNSS jammer classification on tiny models. This matters for anti-jam receivers and compact CRPA antennas that must pick a mitigation method before tracking loops fail.","discovery_kind":"extension","skeptic_critique":{"model":"grok-4.5","headline":"Generalization claim rests on single-interferer synthetic training plus label remapping; absolute gains stay modest and unquantified for variance.","rationale":"The reader correctly isolated the weakest link: single-interferer synthetic training plus taxonomy remapping as the generalization story. The paper’s own numbers already show only modest absolute gains and a large drop on the short-window public set; without error bars or an unmapped/multi-interferer check, those gains cannot be treated as robust. This does not overturn the engineering usefulness of the virtual-ANF features, but it keeps the paper accept-shaped only conditionally on clearer generalization evidence and artifact release—exactly the reader’s CONDITIONAL verdict. No stronger internal inconsistency was found; the math and feature construction are standard and the directional complementarity is plausible from Figs. 1–2 and Table 2.","tokens_in":11586,"tokens_out":554,"duration_ms":10655,"concrete_test":"Retrain the identical z2 GBDT on the EDGE synthetic split, then evaluate ANF+FFT vs FFT-only on (a) the unmapped multi-class DARCY labels without the seven-family collapse and (b) a controlled two-interferer overlay of the EDGE hardware recordings; report per-family accuracies with 5-seed means and 95 % CIs. If the ANF+FFT lift disappears or reverses on either set, the generalization half of the strongest claim fails.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim (Table 1) is that ANF+FFT consistently beats FFT-only for compact GBDTs across three datasets. That claim is load-bearing on the assumption that training solely on the authors’ matched single-interferer EDGE simulator (N=8192, Fs=60 MHz) and then evaluating on EDGE hardware recordings plus a heavily filtered/mapped DARCY subset is enough to establish real generalization. Section II.1 explicitly restricts the signal model to one interferer; DARCY is reduced to ~3500 short (N=1024) snapshots whose original labels are forcibly mapped onto the seven-family taxonomy. Absolute lifts are small (FFT-only 75.52/77.61/57.29 % → ANF+FFT 79.46/79.92/59.12 %), no confidence intervals or multiple-seed statistics appear, and the hardest public set remains near 59 %. If the mapped short-window regime or multi-interferer scenes erase the complementary value of the seven ANF statistics, the “consistently improves … across … datasets” claim does not hold at the strength asserted.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The paper addresses real-time pre-correlation GNSS RFI classification under embedded compute/memory limits. It introduces a compact feature vector derived from a virtual (estimation-only) one-pole complex ANF with NLMS frequency tracking—statistics of unwrapped pole angle including mean, variance, least-squares slope, and first/second-difference variances—and combines these with conventional FFT/STFT spectral descriptors. Compact gradient-boosted decision trees (XGBoost) are trained on a large matched synthetic EDGE Microwave dataset and evaluated on held-out synthetic data, EDGE hardware recordings, and a label-mapped subset of the public DARCY dataset under a seven-family taxonomy. Tables 1–2 and a complexity–accuracy bubble plot (Fig. 3) report that ANF+FFT consistently outperforms ANF-only and FFT-only inputs for family-level accuracy, with mild hyperparameter sensitivity (Fig. 4). The work targets deployment on a compact CRPA (HEDGE8008).","tokens_in":11881,"tokens_out":1293,"duration_ms":33872,"significance":"If the complementary value of the virtual-ANF statistics holds under broader conditions, the contribution is practically useful: inexpensive, fixed-point-friendly expert features that improve compact GBDT accuracy without spectrogram DNNs, which matters for pre-correlation mitigation selection on CRPAs and embedded receivers. Strengths include a clear signal model and ANF derivation (Eqs. 2–12), multi-dataset evaluation including a public corpus, an ablation over ka, µ, and STFT size/hop, and an explicit complexity–accuracy tradeoff rather than accuracy alone. The engineering framing (single-interferer families, lightweight models) is appropriate to the stated deployment goal.","major_comments":[{"comment":"Abstract, contributions bullet 3, and §I promise quantified resource utilization on the EDGE Microwave HEDGE8008 CRPA for the proposed models and literature baselines. §IV.3 and Fig. 3 report only estimated memory footprint and operation counts for GBDT size sweeps, not measured latency, memory, or FPGA/SoC utilization on HEDGE8008. Either add the promised platform measurements (with baselines) or narrow the claim to offline complexity estimates; as written the deployment claim is unsupported in the body.","section":"Abstract; §I; §IV.3; Fig. 3"},{"comment":"The central generalization claim (Table 1: ANF+FFT best on synthetic, EDGE-recorded, and DARCY) rests on training solely on the authors’ matched single-interferer simulator (§II.1 explicitly assumes one interferer; N=8192, Fs=60 MHz) and on forcibly mapping/filtering DARCY to ~3500 short (N=1024) snapshots under the seven-family taxonomy. Absolute lifts are modest (e.g., FFT-only 75.52→79.46% synthetic; DARCY remains ~59%). No confidence intervals, repeated-seed variance, or multi-interferer stress tests are reported. The paper should quantify uncertainty on Table 1–2 accuracies and state clearly that multi-interferer and multi-vendor front-end/AGC generalization are untested, or add such experiments.","section":"§II.1; §IV.2; Table 1; Table 2"},{"comment":"Table 2 shows that ANF+FFT does not uniformly help every family (e.g., NB Modulated 92.19% vs FFT-only 92.26%; Broadband Pulse essentially tied). The narrative that ANF improves “both narrowband and broadband non-stationary RFI” should be aligned with per-family results—highlighting where complementarity is real (CW Pulsed, Chirp, CW Static) versus neutral—so the complementary-feature claim is not overstated relative to the data.","section":"§IV.2; Table 2; Abstract"}],"minor_comments":[{"comment":"Figs. 1–2 are dense multi-panel plots; axis labels and which panel is which jammer type are hard to parse in text form. Add a panel legend mapping (a)–(l) to family names and ensure θz and STFT color scales are readable.","section":"Figs. 1–2"},{"comment":"Notation: z0[n], θz, ϕ, and ϕc are introduced clearly in §III, but Fs appears in the frequency conversion without being defined in that section; define sampling rate when first used.","section":"§III"},{"comment":"The classifier is repeatedly called “z2” in Table 1–2 captions without definition in the main text; define the GBDT configuration (depth, estimators, etc.) once.","section":"Table 1; Table 2"},{"comment":"Typos/orthography: “avirtual” spacing, “T¨ urkiye”/encoding artifacts in biographies, “quite case” → “quiet case” in §IV.1, and inconsistent “EDGE Microwave” spacing.","section":"Throughout; §IV.1"},{"comment":"Related work cites strong recent GNSS RFI ML papers; a short explicit comparison of feature dimensionality and claimed embedded feasibility versus van der Merwe et al. (2024) and Mehr & Dovis (2025) would help readers place the contribution.","section":"§I; References"}],"recommendation":"major_revision","confidential_remarks":"Solid applied DSP/ML engineering paper with a plausible lightweight feature idea; not a fundamental methods breakthrough. The missing HEDGE8008 measurements are the clearest fix-or-retract item given how prominently they are advertised. Fit for an applications-oriented signal-processing or navigation venue after revision; reject only if the journal expects strong multi-condition statistical generalization that the current single-interferer, single-simulator training setup cannot support."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"Punchline: they freeze a classical one-pole ANF in estimation-only mode, turn the unwrapped pole-angle trajectory into seven cheap statistics, glue them to FFT features, and get consistent family-level lifts on compact GBDTs. That is real, useful, and not oversold in the body.\n\nWhat is new is the specific feature vector (energy, mean/var of φ, slope, first/second-diff variances, median-filtered diff variance) and the empirical case that it complements STFT where STFT is weak—pulsed CW, some chirps, matched BPSK—without blowing the embedded budget. The math is standard Borio-style NLMS; no invention theater. Tables 1–2 and the complexity bubble plot are the core evidence: ANF alone is weak, FFT alone is better, ANF+FFT wins across their synthetic split, their hardware recordings, and a mapped DARCY subset, and the hybrid saturates later as model size grows. Hyperparameter ablation is mild and reported. Embedded motivation is clear.\n\nSoft spots, in proportion: gains are modest (roughly +4 points on synthetic, smaller on DARCY’s ~59%). No error bars or multi-seed variance. Signal model is single-interferer by design. External labels are remapped onto their seven families, and training is on a matched EDGE simulator—so “generalization” is same-lab hardware plus a short-window public subset, not multi-vendor AGC chaos. Abstract promises HEDGE8008 resource numbers; the text is stronger on accuracy/complexity than on full platform utilization detail. None of that sinks the central claim; it just caps how far you should push it.\n\nWho it is for: people building pre-correlation classifiers on CRPAs and tight SoCs who already live in expert-feature + GBDT land. Not a theory paper. Citation pattern is appropriate (Borio ANF, recent GNSS ML). I would send it to peer review; it is coherent, checkable DSP, and honest about scope. Engage if you care about lightweight GNSS RFI features; skip if you need multi-interferer theory or large absolute accuracy jumps.","headline":"Solid incremental engineering: virtual-ANF angle stats plus FFT give modest, consistent gains for compact GBDTs; generalization story is honest but thin.","tokens_in":12512,"tokens_out":536,"would_cite":false,"duration_ms":22138,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"Virtual adaptive-notch frequency statistics plus FFT features improve compact GNSS interference classification under real-time embedded budgets.","keywords":["GNSS","RFI classification","adaptive notch filter","pre-correlation","gradient-boosted decision trees","embedded CRPA","spectral features","instantaneous frequency"],"falsifier":"Retrain the same GBDT on the authors’ synthetic set, then evaluate ANF+FFT versus FFT-only on independent multi-interferer field captures from different front-ends and sampling rates; if the accuracy lift disappears or reverses, the central complementarity claim fails.","tokens_in":12414,"feed_emoji":"📡","tokens_out":822,"duration_ms":16792,"temperature":0.7,"pith_summary":"GNSS receivers in safety-critical settings need to know what kind of radio interference they face so they can pick the right countermeasure before tracking loops fail. Rule-based detectors do not cover the full range of jammers, and large neural nets are too heavy for compact embedded platforms. This paper shows that a cheap “virtual” adaptive notch filter—run only to estimate instantaneous frequency, not to suppress interference—yields a handful of statistics that separate narrowband, chirp, pulsed, and broadband jammers. When those statistics are concatenated with ordinary FFT spectral features and fed to small gradient-boosted trees, family-level accuracy rises over either feature set alone on synthetic data, lab recordings, and a public short-snapshot set, while staying light enough for a compact CRPA board.","feed_headline":"Cheap notch-filter stats lift GNSS jammer classification","feed_subtitle":"Virtual ANF features plus FFT beat either alone on compact trees for real-time receivers","key_machinery":"The virtual ANF feature vector: seven cheap statistics (energy, mean and variance of unwrapped pole angle, least-squares slope, and first/second-difference and median-filtered difference variances) computed from the NLMS frequency tracker of a one-pole complex notch run in estimation-only mode on each baseband window.","core_discovery":"Lightweight statistics taken from the instantaneous-frequency trajectory of a virtual adaptive notch filter complement conventional FFT descriptors; together they raise pre-correlation GNSS RFI family classification accuracy for compact gradient-boosted decision trees across synthetic, hardware-recorded, and public datasets, without requiring heavy neural networks.","pith_inferences":["The same pole-trajectory statistics could serve as cheap latent inputs for online or weakly-supervised adaptation when labeled field data are scarce.","Because the ANF already models a single tone, multi-tone or dense multi-jammer scenes will likely need an explicit multi-pole or bank extension before the accuracy lift holds.","Fixed-point NLMS ANF features map naturally onto existing FPGA notch-filter IP, so the method can reuse silicon already present for mitigation."],"forward_implications":["Embedded CRPA and receiver modules can add the virtual-ANF stats with negligible extra arithmetic and still raise jammer-family accuracy.","For a fixed memory or cycle budget, ANF+FFT reaches a target accuracy with a smaller tree ensemble than FFT alone.","Mitigation chains can be switched or parameterized earlier, before loss of lock, using only pre-correlation baseband windows.","Hyperparameter sensitivity of both ANF and FFT is modest (~±2 % accuracy), easing calibration on new platforms."],"fun_headline_variants":["Virtual ANF stats plus FFT lift GNSS RFI classification on compact trees","Notch-filter frequency features boost pre-correlation jammer family accuracy","Lightweight ANF trajectory stats complement FFT for real-time GNSS RFI","Compact GBDT classifies GNSS interference with virtual notch-filter features","ANF instantaneous-frequency stats raise GNSS jammer accuracy without NNs"],"cache_read_input_tokens":128,"weakest_assumption_plain":"Training only on a matched front-end simulator and testing under a single-interferer, seven-family taxonomy is enough to claim the method will generalize to real multi-vendor receivers, AGC behavior, and multi-jammer scenes.","fun_headline_variants_meta":{"raw":{"variants":["Virtual ANF stats plus FFT lift GNSS RFI classification on compact trees","Notch-filter frequency features boost pre-correlation jammer family accuracy","Lightweight ANF trajectory stats complement FFT for real-time GNSS RFI","Compact GBDT classifies GNSS interference with virtual notch-filter features","ANF instantaneous-frequency stats raise GNSS jammer accuracy without NNs"]},"model":"grok-4.5","effort":"low","cost_usd":0.003285,"raw_usage":{"total_tokens":1140,"prompt_tokens":780,"num_sources_used":0,"completion_tokens":78,"cost_in_usd_ticks":32848000,"prompt_tokens_details":{"text_tokens":780,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":282,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":780,"tokens_out":78,"duration_ms":6436,"temperature":1.0,"reasoning_tokens":282,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-31T08:24:00.602127+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Retrain the same GBDT on the authors’ synthetic set, then evaluate ANF+FFT versus FFT-only on independent multi-interferer field captures from different front-ends and sampling rates; if the accuracy lift disappears or reverses, the central complementarity claim fails.","supporting_citations":[],"review_version":1}