{"id":"a12c0254-147f-400d-b580-a4c23abaa865","arxiv_id":"2506.16099","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"BLADE uses Savitzky-Golay filtering, prominence-based peak detection, and gradient analysis to classify 124 CNEOS bolide light curves into fragmentation behavior categories.","lead":"This paper introduces BLADE, an automated software pipeline that sorts fireball light curves from NASA's CNEOS database into categories such as airburst, single peak, and continuous or discrete fragmentation. A generalist might read it as a demonstration of how standard signal processing can be applied to a growing archive of asteroid impact events, relevant to planetary defense and atmospheric monitoring.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No independent validation of peak-physics correspondence: the only thresholds are tuned on and applied to the same 124 events, so 'reliably distinguishes' is untested against any ground truth; CNEOS digitization and smoothing could be classifying artifacts.","rationale":"The reader's weakest assumption identifies exactly the load-bearing concern: the physical meaning of detected peaks. I refine it into a concrete validation gap with a circular-tuning component: thresholds were tuned on the same dataset that is then classified, and no independent ground-truth comparison exists anywhere in the paper. The Chelyabinsk counterexample (Section 4) is internal evidence that the CNEOS light curves can miss real fragmentation structure, which is particularly damaging because BLADE's classification rests entirely on peak structure. I also flag the unquantified PDF digitization step as a source of peak distortion that is directly relevant to thresholds defined in normalized amplitude and sample counts. The paper is honest about limitations and the framework is deterministic and well described, so this is not a rejection; the abstract's 'reliably distinguishes' claim simply outruns the evidence. The proposed independent-validation and re-digitization tests would settle whether the concern lands. Verdict: CONDITIONAL, requiring external validation and digitization-error assessment before the central claim can be accepted.","tokens_in":23107,"tokens_out":4123,"duration_ms":40098,"concrete_test":"Select 20-30 of the 124 events that have independent fragmentation evidence (e.g., Chelyabinsk video records, published all-sky camera observations, or infrasound-derived energy-deposition profiles). Run BLADE exactly as specified and compare the resulting classifications and peak counts against the independent fragmentation records; compute agreement statistics. If agreement is low, the thresholds are not capturing physical events. In parallel, re-digitize 10 light curves from the CNEOS PDFs with a different operator or tool and rerun BLADE to measure how much peak counts and class labels change with digitization; if labels change for more than 10% of the re-digitized events, digitization error is a dominant uncertainty in the claimed reliability.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim requires that peaks in the digitized, Savitzky-Golay smoothed, normalized CNEOS light curves correspond to physical fragmentation or energy-release events, and that BLADE's categorical labels are reliable. The paper provides no test of this correspondence. All adjustable parameters were selected on the same data they later classify: prominence=0.10, height=0.10, and distance=5 were 'empirically determined through a systematic examination of representative fragmentation and airburst profiles' (Section 2.2.2), and the gradient threshold of 0.5 was selected based on 'sensitivity analyses of representative bolide events within the CNEOS dataset' (Section 2.2.3). No held-out set, no independent reference, and no comparison to video, infrasound, or physics-based fragmentation models is presented; Section 5 lists 'validation and benchmarking' only as future work. The paper itself supplies the strongest counterexample: Section 4 notes that the CNEOS Chelyabinsk light curve shows a single dominant peak while video recordings show complex fragmentation (Brown et al., 2013), implying the sensor or data product can severely distort peak structure. Since the light curves are distributed only as PDFs and must be manually digitized (Section 2.1), digitization error is an additional uncontrolled source of peak-shape distortion, yet no reproducibility check is reported. Consequently, the abstract's 'reliably distinguishes distinct bolide behaviors' is currently an untested assertion about a circular tuning loop, not a demonstrated property of the framework.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces BLADE, an automated pipeline for classifying light curves in the CNEOS fireball database. The pipeline digitizes PDF light curves, applies a Savitzky-Golay filter with noise-adaptive window parameters, normalizes the signal, detects peaks via SciPy's find_peaks with thresholds on prominence, height, and separation, and computes gradients. Based on the number and temporal spacing of detected peaks and on the maximum gradient, each event is assigned to one of five categories: continuous fragmentation, discrete fragmentation, airburst, single peak, or no significant peaks. The authors apply BLADE to 124 CNEOS fireballs with infrasound detections and report that 65.9% are classified as discrete fragmentation, 24.6% as single peak, 7.1% as airburst, and 2.4% as continuous fragmentation. They also demonstrate altitude extrapolation for events with velocity metadata and discuss the relationship of their categories to energy deposition and infrasound. The central claim, stated in the abstract, is that BLADE 'reliably distinguishes distinct bolide behaviors.'","tokens_in":23426,"tokens_out":4324,"duration_ms":53034,"significance":"If the reliability claim could be supported, BLADE would fill a practical niche: it offers a deterministic, scalable, and openly specified procedure for turning a large, underused public dataset (CNEOS light curves) into categorical labels for fragmentation behavior. The pipeline is clearly described, the choice of Savitzky-Golay filtering and prominence-based peak detection is sensible for noisy, heterogeneous time series, and the authors provide a supplementary CSV of their classifications, which will be useful regardless of the label semantics. The paper explicitly positions the work as a foundation, and the code-level details (parameter choices, library calls) are given in enough detail to allow independent implementation. The main significance is therefore conditional: the paper is a promising methodological blueprint, but its headline claim of reliability is not yet supported by the evidence in the manuscript.","major_comments":[{"comment":"The central claim that BLADE 'reliably distinguishes distinct bolide behaviors' is not supported by any validation experiment. All detection thresholds (prominence≥0.10, height≥0.10, distance=5, gradient>0.5) were set after 'systematic examination' of the same CNEOS dataset on which performance is then demonstrated (Section 2.2.2, Section 2.2.3). There is no held-out set, no comparison to independent physical ground truth (video, radar, infrasound, or physics-based fragmentation simulations), and no human-expert agreement measure. Section 5 lists 'Validation and benchmarking' only as future work. Given that the categories are the paper's main output, this is a load-bearing gap. The authors should add a validation section with external references (e.g., well-documented events like Chelyabinsk, or synthetic light curves with known fragmentation sequences) or at minimum a cross-validation scheme and inter-labeler agreement analysis.","section":"Abstract and §2.2.2–§2.2.3 and §5"},{"comment":"The paper does not quantify digitization error, although the input data are PDF light curves that must be manually digitized with a web-based tool before peak detection. Because the entire classification hinges on the number, position, and prominence of peaks, uncontrolled digitization error can change labels. No reproducibility check (e.g., re-digitizing a subset of light curves multiple times and comparing the resulting classifications) is reported. Without this, the phrase 'high-fidelity framework' in the abstract and the claim of 'reliable' classification are not justified. The authors should add a digitization-uncertainty analysis and propagate the resulting variability through the peak detection and classification steps.","section":"§2.1 and §3"},{"comment":"The manuscript itself provides a direct counterexample to the assumption that peaks in CNEOS light curves correspond to physical fragmentation events: the Chelyabinsk light curve shows a single dominant peak, while video recordings show complex fragmentation. This is acknowledged in Section 4 but is not resolved or used to qualify the classification labels. As written, BLADE's categories ('discrete fragmentation', 'continuous fragmentation', 'airburst') imply a physical interpretation that the data product may not support. The authors should either (i) restrict the claims to morphological classification of the digitized light curves, or (ii) provide evidence for a subset events with independent observations that the detected peak structure indeed tracks physical fragmentation. This distinction is essential for the paper's advertised use in energy-deposition and infrasound studies.","section":"§4, Chelyabinsk discussion"},{"comment":"The boundary between 'discrete fragmentation' and 'continuous fragmentation' rests on a peak separation threshold of '10–15 samples' that is not derived from any physical timescale and is not accompanied by a sensitivity analysis. The paper reports that nine of 124 events are ambiguous between these two categories (Section 3.7), yet the threshold's choice is presented as a given. Furthermore, the distinction between 'single peak' and 'airburst' is acknowledged to be fuzzy (Section 3.6, Section 4), and the aggregate statistics (Figure 9) treat these as separate categories. If the method is to be a foundation for future physics studies, the classification criteria need a more explicit and less arbitrary basis, or at least a quantitative sensitivity study showing how the category fractions vary with the thresholds.","section":"§2.2.4 and §3.7"}],"minor_comments":[{"comment":"The text in §3.6 describes the event of 2022-02-07 20:06:25 UTC, but the Figure 8 caption reads 'Light curve analysis for the 2018-06-21 01:16:20 UTC event.' One of these is a typo; please correct it so that the event identifier is consistent.","section":"Fig. 8 caption and §3.6"},{"comment":"The sentence 'While not pertinent to BLADE, we plotted the distribution...' appears to contain a typo; it should likely say 'While not central to BLADE' or 'While not the primary focus of BLADE.'","section":"§3.7 first sentence"},{"comment":"The noise-level bins (σΔI ≤ 0.5, 0.5 < σΔI ≤ 1.0, σΔI > 1.0) and the corresponding window lengths and polynomial orders are described in prose but not summarized in a table; a small table would aid reproducibility.","section":"§2.2.1, noise-adaptive SG parameters"},{"comment":"The text explains that the 0.5 gradient threshold can be interpreted either per second or per sample depending on whether a time vector is supplied to np.gradient, but it never states which convention was used for the 124-event analysis. Please specify the actual implementation.","section":"§2.2.3, gradient threshold units"},{"comment":"The statement that 'BLADE source code is not publicly available at this time' is in tension with the paper's reproducibility emphasis; consider releasing at least a minimal reference implementation or a containerized version, or clarify why the code cannot be shared.","section":"Data Availability"},{"comment":"'This dynamic adjustment optimizess the balance' should read 'optimizes.'","section":"§2.2.1, typo"}],"recommendation":"major_revision","confidential_remarks":"The paper was already accepted in AJ per the header, so the referee report should be understood as a post-acceptance review. The core methodological contribution is real, but the abstract overstates reliability. I would urge the editor to require the authors to temper the abstract and add either validation or strong caveats before publication; the current version risks being cited as proof of physical classification that is not demonstrated."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe short version: BLADE is a straightforward, deterministic pipeline—Savitzky-Golay filtering, peak prominence, gradient thresholds—applied to 124 CNEOS fireball light curves. That is the new part: a public, reproducible set of category labels (airburst, discrete/continuous fragmentation, single peak) for a sample that overlaps infrasound-detected events. The paper is honest about what it does not do, and the supplementary CSV makes the output usable. If you work on bolides or infrasound correlation, this is worth a look.\n\nWhat it does well: the method section is unusually explicit. The equations, parameter choices, and even the exact SciPy calls are given. The adaptive noise-based filtering is a nice touch, and the authors flag that the CNEOS PDFs require manual digitization, which is a real uncontrolled source of error. The discussion of ambiguity (nine events) is measured.\n\nWhere it is soft: the central claim—'reliably distinguishes distinct bolide behaviors'—is not actually tested. All thresholds were chosen by looking at representative events from the same 124, then applied back to those events. There is no held-out set, no comparison to video, infrasound, or physics-based fragmentation models, and no cross-validation. The paper's own Chelyabinsk example undercuts peak-physics correspondence: the CNEOS light curve shows one peak while video shows complex fragmentation. The authors list validation as future work, which is fine for a preliminary methods paper, but the abstract's 'reliably' is doing too much work. Also, code is not released, so 'reproducible' rests on the PDF digitization step being done exactly as they did it, which is not fully specified.\n\nProportionally, these are fixable problems rather than fatal ones. The categories are heuristic, but the framework is transparent and adaptable. The absence of ground truth is the main limitation, and the paper acknowledges most of it.\n\nMy take: this is a solid desk-review pass. I would send it to a referee with a note that the abstract should be toned down and validation against at least a handful of well-observed events (video or infrasound) would strengthen the paper. The classification labels themselves are a useful community resource, even if the precision of individual labels remains uncertain.","headline":"A useful, clearly written classifier for CNEOS bolide light curves, but the reliability claim is untested; still deserves a serious referee.","tokens_in":23916,"tokens_out":1622,"would_cite":true,"duration_ms":19535,"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":"BLADE classifies fireball light curves into four fragmentation modes automatically.","keywords":["bolides","fireballs","light curve classification","fragmentation","CNEOS database","Savitzky-Golay filter","infrasound","planetary defense"],"falsifier":"Run BLADE on fireballs that also have independent high-cadence video trajectory and fragmentation reconstructions, then check whether each detected peak appears as a fragmentation event at the same time and altitude in the video; any systematic mismatch between detected peaks and observed fragmentation times, or a failure to reproduce Chelyabinsk's complex fragmentation, would falsify the peak-to-event mapping.","tokens_in":22934,"feed_emoji":"☄️","tokens_out":6815,"duration_ms":70062,"temperature":0.7,"pith_summary":"The paper introduces BLADE, an automated pipeline that classifies bolide light curves from the NASA CNEOS fireball database. It claims that smoothing, peak detection, and gradient analysis reliably separate airbursts, discrete fragmentation, continuous fragmentation, and single-peak events without manual tuning. Applied to 124 fireballs with known infrasound detections, the method completes the whole analysis in minutes, and only nine cases were ambiguous between two fragmentation classes. This matters because objective, reproducible classification of large light-curve datasets is needed to connect observed brightness to energy deposition and acoustic signatures, supporting planetary defense and global atmospheric monitoring.","feed_headline":"Automated pipeline sorts fireball light curves into four classes","feed_subtitle":"Brightness-peak pipeline labels airbursts, discrete and continuous fragmentation, and single-peak events","key_machinery":"The load-bearing object is the sequence of normalized light-curve peaks extracted by a three-step procedure: adaptive Savitzky-Golay filtering, which sets window length and polynomial order from the standard deviation of the first-difference series (the estimated noise floor); prominence-based peak detection with fixed thresholds of 0.10 normalized intensity and a five-sample minimum separation; and gradient analysis with a 0.5 threshold to identify steep rises. The spacing of the detected peaks (closer or farther than roughly ten to fifteen samples) then separates continuous from discrete fragmentation, while a single steep-rise peak is classified as an airburst or single-peak event.","core_discovery":"The paper argues that a fully signal-driven pipeline, with no physical assumptions about meteoroid structure or ablation, can assign most CNEOS light curves to physically meaningful classes based solely on the number, spacing, prominence, and steepness of brightness peaks. On a subset of 124 bolides with accompanying infrasound detections, the pipeline labels each event as airburst, discrete fragmentation, continuous fragmentation, or single peak; 65.9% fall in discrete fragmentation, 24.6% single peak, 7.1% airburst, and 2.4% continuous fragmentation. Only nine events were ambiguous, and in every ambiguous case fragmentation was still correctly identified as the primary process. The paper also shows that detected peaks can be anchored to approximate altitudes when velocity-vector metadata exist, and that the analysis takes only minutes for the full 124-event set.","pith_inferences":["A natural validation the paper leaves open is to compare BLADE's class labels against independent video-observed fragmentation sequences; Chelyabinsk already shows that the CNEOS light curve can hide real structure, so this comparison would test whether the other 123 labels are physically meaningful.","Because the thresholds (0.10 prominence, 0.5 gradient, five-sample separation) were chosen empirically on this 124-event sample, porting BLADE to differently sampled sensors will likely require rescaling thresholds relative to noise rather than using the fixed values.","If the peak-to-fragmentation mapping is correct, BLADE's derived peak altitudes should agree with infrasound-derived source heights for the same events; that cross-modal correlation is the paper's stated future direction and would provide an independent check.","The 124-event subset was deliberately built from events with published infrasound detections, so the reported proportions (65.9% discrete fragmentation, for instance) describe that multimodal subset, not necessarily the global bolide population."],"forward_implications":["If BLADE works as described, the full archive of roughly 850 light curves can be classified in minutes without manual re-tuning, yielding a uniform fragmentation census.","The automatic class labels can be correlated with altitude, entry angle, velocity, and impact energy to test physical trends, such as airbursts preferring steeper entry trajectories.","Because the method is deterministic and signal-only, the same categories transfer to other high-cadence space-optical datasets, including Geostationary Lightning Mapper bolide observations.","The optional altitude extrapolation converts detected peaks into approximate heights of energy release for the events with velocity metadata, providing input for infrasound-yield and dynamic-pressure modeling.","Only nine of 124 events were ambiguous, always between continuous and discrete fragmentation, and in all cases fragmentation remained the primary classification."],"supporting_citations":[{"why":"Supplies the Savitzky-Golay smoothing operator that removes noise while preserving peak features.","marker":"Savitzky and Golay, 1964"},{"why":"Supplies the first-difference noise estimate and numerical gradient definitions used to set filter parameters and identify steep rises.","marker":"Press et al., 2007"},{"why":"Provides the NumPy routine used to compute the noise floor as the standard deviation of the first differences.","marker":"Harris et al., 2020"},{"why":"Provides the prominence-based find_peaks routine that detects fragmentation peaks in the smoothed light curve.","marker":"Virtanen et al., 2020"},{"why":"One of the infrasound catalogs used to select the 124-event dataset with coincident acoustic detections.","marker":"Ens et al., 2012"},{"why":"Documents the Chelyabinsk light-curve limitation used to temper the peak-to-fragmentation interpretation.","marker":"Brown et al., 2013"},{"why":"Establishes the empirical energy relationship behind the CNEOS impact-energy values used in the analysis.","marker":"Brown et al., 2002"}],"fun_headline_variants":["Fireball light curves auto-classified into four classes","No-assumption pipeline sorts bolide light curves into four types","Signal-only tool labels airbursts and fragmentation events","Automated sorter identifies airburst, fragmentation, single-peak"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The classification assumes a peak in the digitized, smoothed light curve equals a physical fragmentation or energy-release event, an assumption the paper itself weakens by noting that Chelyabinsk shows a single peak while video shows complex fragmentation.","fun_headline_variants_meta":{"raw":{"variants":["Fireball light curves auto-classified into four classes","No-assumption pipeline sorts bolide light curves into four types","Signal-only tool labels airbursts and fragmentation events","Automated sorter identifies airburst, fragmentation, single-peak"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001047,"raw_usage":{"total_tokens":4417,"prompt_tokens":978,"completion_tokens":3439,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":594,"completion_tokens_details":{"reasoning_tokens":3371}},"tokens_in":594,"tokens_out":3439,"duration_ms":27683,"temperature":1.0,"reasoning_tokens":3371,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T23:44:13.742443+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run BLADE on fireballs that also have independent high-cadence video trajectory and fragmentation reconstructions, then check whether each detected peak appears as a fragmentation event at the same time and altitude in the video; any systematic mismatch between detected peaks and observed fragmentation times, or a failure to reproduce Chelyabinsk's complex fragmentation, would falsify the peak-to-event mapping.","supporting_citations":[],"review_version":1}