{"id":"47fd17a7-bf92-496f-b7ee-491cff8337bb","arxiv_id":"2507.10307","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"A blob-detection pipeline can seed gamma-ray source locations much faster than blind likelihood searches, but the paper's accuracy and speedup claims are inconsistent with its own results.","lead":"This paper applies an image-processing algorithm that finds bright spots in gamma-ray sky maps and feeds them as starting guesses to a detailed source-fitting routine. The authors report large speedups in seeding sources, but the performance numbers in the paper conflict with each other and with the reported recovery rates.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The paper's own recovery counts contradict its accuracy claims: only 1 of 33 large extended sources is recovered, while §4 claims >95% accuracy for extended sources, so the central claim of accurate source seeding is not supported.","rationale":"The paper presents a plausible image-processing approach and honestly reports per-category recovery counts; those counts are the most informative evidence available. The decisive problem is that the claims in the abstract and conclusion are contradicted by those counts. Large extended sources are the hardest case for seeded searches because the DoG bandpass filter suppresses low-frequency structure; the text itself notes they are removed. With only 1/33 recovered, the method cannot be said to 'seed sources accurately' for the full source population it is meant to help find. This is an internal inconsistency, not a dispute with external consensus, so it is load-bearing regardless of whether parameters generalize. I do not see a need to charge the authors with anything beyond reporting inconsistencies; the fix is to report the confusion matrix honestly and reword the claims, or to extend the pipeline so large extended sources are recoverable. The reader's weakest assumption about in-sample tuning is a real secondary concern, but the arithmetic contradiction is more decisive, so my agreement is partial.","tokens_in":5307,"tokens_out":4393,"duration_ms":52420,"concrete_test":"Re-derive the aggregate confusion matrix from the paper's own counts: count a seed as recovered if it matches a simulated source within the paper's matching criterion and split by the three size bins. If the large-extended bin is 1/33 and the overall extended recovery is 25/58, then the §4 '>95% extended-source accuracy' claim is false regardless of parameter generalization. A stronger external check would rerun the pipeline on held-out simulations with fixed parameters and report recovery per size bin.","verdict_should_be":"UNCHANGED","load_bearing_attack":"In Section 3 the reported recovery is 65/66 point sources, 24/25 small extended sources, and 1/33 large (>0.80 deg) extended sources. The Section 4 conclusion that the algorithm achieves >95% accuracy for extended sources is arithmetically incompatible with these numbers: 25 of 58 extended sources is 43%. This matters because the DoG preprocessing in Section 2.3 is described as removing large extended sources, so the failure is structural rather than an incidental threshold miss. The central advertised result — accurate seeding of sources for the HAWC likelihood pipeline — is therefore not established for the source class that is hardest for the existing blind search. In addition, the speed claim is not controlled: 'less than 26 hrs' for the image pipeline over 60 simulation sets is compared with 'over 2 months' for the conventional search without stating the endpoint, hardware, or whether identical ROIs and simulations were used, and the abstract's '300 times faster' appears nowhere in Section 3, which reports 60×.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript describes an image-processing pipeline for seeding gamma-ray source candidates in HAWC significance maps. The pipeline uses a Difference-of-Gaussians (DoG) filter, a blob detector, and intensity thresholding to produce a list of source seeds that are then passed to a likelihood-based fitting tool. Performance is reported on 60 simulated maps containing 123 sources, with claims of a large speedup over the conventional blind search. The paper concludes that the algorithm achieves >98% accuracy for point sources and >95% accuracy for extended sources, with a computational speedup of about 98%. The central claim is that this pipeline can accelerate and improve HAWC source searches.","tokens_in":5488,"tokens_out":3008,"duration_ms":31914,"significance":"If the claims were supported, the pipeline would be a useful complement to the computationally expensive likelihood searches used by HAWC and similar observatories. The paper addresses a real operational problem and proposes a concrete, potentially transferable method. However, the reported results are undermined by internal inconsistencies between the abstract, the performance section, and the conclusion, and by the absence of any validation on independent or real data. As written, the paper does not establish its central advertised results for extended sources or the '300 times faster' speedup, so its practical significance cannot be assessed reliably.","major_comments":[{"comment":"The recovery numbers in Section 3 do not support the accuracy claims in Section 4. Section 3 reports recovery of 65/66 point sources, 24/25 small extended sources, and 1/33 large extended (>0.80 deg) sources. For extended sources, this is 25 of 58, or about 43%, not '>95% accuracy for extended sources' as claimed in Section 4. The failure on large extended sources is not incidental: Section 2.3 states that 'large extended sources are removed' by the DoG filter. The central claim of accurate source seeding is therefore not supported for the source class that is most challenging for the existing blind search, and the conclusion is arithmetically incompatible with the paper's own data.","section":"Section 3 and Section 4"},{"comment":"The abstract claims the pipeline 'seeds sources accurately up to 300 times faster than the current HAWC source search pipeline,' but Section 3 reports a '60× speedup' and says nothing about 300×. The number 300 appears nowhere in the text or tables. Moreover, the timing comparison is not controlled: 'less than 26 hrs' for the image pipeline over 60 simulation sets is compared with 'over 2 months' for the conventional search without specifying the computational endpoint, hardware, region of interest, or whether the same simulations and analysis settings were used. The speedup factor and the '98.33% performance gain' should be derived from a clearly defined, matched comparison.","section":"Abstract and Section 3"},{"comment":"The evaluation is in-sample and the key parameters are hand-picked. Section 2.3 states 'A smearing radius of 0.20 is used in the simulation plots shown in the proceeding,' and Section 2.1 sets a 5-sigma detection threshold, but no sensitivity study, cross-validation, or hold-out dataset is presented. Recovery rates and speedups are measured on the same 60 simulations used to motivate these choices. Without testing on independent simulations or on real HAWC maps, the reported accuracy and speedup cannot be interpreted as general performance measures; they are partly a product of the chosen thresholds.","section":"Section 2.3 and Section 3"},{"comment":"The paragraph following Figure 3 states that 'for the simulation dataset used, the image processing pipeline is fully capable of recovering the simulated sources.' This directly contradicts the quantitative recovery counts in the preceding paragraph, where only 1 of 33 large extended sources is recovered. The figure appears to show only a subset of the simulated sources, and the text should be corrected to state the overall recovery fractions, including the poor performance on large extended sources, rather than a qualitative claim of full capability.","section":"Section 3 (Figure 3)"}],"minor_comments":[{"comment":"There are several typos: 'likeihood' in the Introduction, 'greather' in Section 2.4, 'the the pixel' in Section 2.4, and 'proceeding' should be 'proceedings' in Section 2.3. These should be corrected.","section":"Throughout"},{"comment":"The classification thresholds are given as '0.150' and '0.150' without units; presumably these are degrees, and the values should be written as '0.15°' for clarity. The blob scale definition is also not defined precisely.","section":"Section 2.4"},{"comment":"The term 'Drips' is introduced without definition or explanation. It is unclear whether this is a name for the seed list, the pipeline output, or something else; please define it.","section":"Section 3"},{"comment":"The initial peak search uses a '≥5σ intensity' threshold, but the relationship between this threshold and the later 5-sigma blob significance threshold in Section 2.4 is not clarified. Having two 5-sigma thresholds with different meanings is confusing.","section":"Section 2.1"},{"comment":"The false-positive rate of '10%' is stated without a denominator or definition. It would help to specify how false positives are counted, e.g., detected seeds not matching any simulated source within a given angular separation.","section":"Section 3"}],"recommendation":"reject","confidential_remarks":"The paper is a collaboration proceedings contribution, and the core idea is potentially useful, but the internal numerical contradictions (abstract vs. Section 3, Section 3 vs. Section 4) are severe enough that the manuscript cannot be accepted in its current form. Even a major revision would need to add controlled timing benchmarks, a hold-out or real-data test, and a substantial rewriting of the conclusions to match the measured recovery rates. As written, the advertised claims are not supported by the paper's own data, which in my view places this beyond a routine major-revision request."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take on arXiv:2507.10307: the idea—use Difference-of-Gaussians blob detection to seed candidate sources for HAWC's likelihood-based search—is reasonable and worth a look, but the paper as written undercuts itself with internal inconsistencies in the headline numbers. The application to HAWC data and the speedup study are new in this specific form, and I'd credit the authors for trying to address a real bottleneck: the iterative likelihood search is slow and misses faint sources. The pipeline's output, passed to a global fitter, is a sensible work-flow.\n\nWhat's actually good: the recovery of 65/66 point sources and 24/25 small extended sources on simulations suggests the method works for compact sources. The false-positive rate of 10% is also reported, which is honest. And the idea of using a smearing radius tied to the PSF is physically motivated.\n\nThe soft spots are large. The conclusion claims \">95% accuracy for extended sources,\" but Section 3 reports recovery of only 1 of 33 large extended sources—that's 3%, not 95%. The DoG step, as the authors themselves note, removes large extended sources; the algorithm is structurally blind to the source class that is hardest for the existing search. The abstract's \"up to 300 times faster\" never appears in Section 3, which reports 60×. The speed comparison is not controlled: \"less than 26 hrs\" over 60 simulation sets versus \"over 2 months\" for the conventional search, with no statement of hardware, endpoints, or whether the same ROIs and simulations were used. The parameters (smearing radius, threshold) are hand-picked and the accuracy is measured on the same simulations used to choose them—no hold-out set, no real-data demonstration, no code or data released.\n\nThese aren't minor blemishes; the central claims of accuracy and speed are either inconsistent or not independently validated. The method may be salvageable, but this paper does not establish it.\n\nWho is this for? Someone working on source seeding pipelines for HAWC, SWGO, or LHAASO might read it for the idea, but they'd have to redo the validation themselves. I would not cite it in its current form.\n\nMy recommendation: a serious editor would not send this to peer review until the numbers are reconciled and the evaluation is made out-of-sample. As-is, I'd desk reject or return for major revision.","headline":"The pipeline idea is sensible, but the paper's own numbers contradict its central claims, so it is not ready for publication as-is.","tokens_in":6046,"tokens_out":2029,"would_cite":false,"duration_ms":21923,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"An image-processing pipeline can seed HAWC gamma-ray sources about 60 times faster than the likelihood search, while recovering nearly all point and compact sources.","keywords":["HAWC observatory","very-high-energy gamma rays","gamma-ray source search","image processing","Difference of Gaussians","blob detection","source seeding","likelihood fitting"],"falsifier":"Run the pipeline with those fixed parameters on a held-out set of simulations, or on a real HAWC significance map whose sources are already known, and count how many injected or catalogued point and compact sources appear in the seed list; if recovery falls well below 65 of 66 point sources and 24 of 25 compact extended sources, or if the wall-clock advantage over the likelihood search disappears in a head-to-head test, the central claim would not transfer.","tokens_in":5073,"feed_emoji":"🔭","tokens_out":12356,"duration_ms":128525,"temperature":0.7,"pith_summary":"This paper is trying to establish that the slowest part of HAWC's blind gamma-ray source search can be replaced with an image-processing step. Instead of fitting a multi-source statistical model to a region directly, which takes days, the pipeline treats the significance map as an image, sharpens it with a Difference-of-Gaussians filter, and pulls out candidate source positions with a blob finder. The candidates are then handed to the usual likelihood fitter for exact localization. On 60 simulated fields the seeds recover 65 of 66 point sources and 24 of 25 compact extended sources, and the image step ran in under 26 CPU-hours total instead of the two months the likelihood search needed. If this holds, HAWC and similar instruments could scan wide regions for new or variable gamma-ray sources much faster and at lower computational cost.","feed_headline":"Image filtering finds HAWC gamma-ray sources 60x faster","feed_subtitle":"Seeds recover 65 of 66 point sources before the statistical fit.","key_machinery":"The machinery is a Difference-of-Gaussians (DoG) spatial band-pass filter combined with a multi-scale blob finder. A Gaussian-blurred version of the normalized significance image is subtracted from the original, so sharp high-frequency peaks survive while broad diffuse emission is removed; the blob finder then scans the residual across increasing blur scales and keeps pixels that are local extrema in several consecutive scales. Fixed parameters do the noise rejection: a 0.20° smearing radius matched to the detector point-spread function, a 3σ threshold on the residual-intensity histogram, a 5σ threshold on the original significance map, and a cut at 10 pixels from the image edge. Detected blobs with scale below 0.15° are labeled point-like and larger ones extended. This machinery reduces a many-parameter likelihood fit to a short candidate list that the fitter only has to refine.","core_discovery":"The central claim is that a standard image-processing stack recovers most gamma-ray sources from HAWC significance maps before any likelihood fitting is done, and does so far faster than the iterative multi-source search now in use. The stack clips the map at -5σ, normalizes it, applies a Difference-of-Gaussians spatial band-pass filter, and runs a multi-scale blob finder that tags local extrema as source seeds. On the 60-simulation benchmark, the pipeline recovered 65 of 66 point sources and 24 of 25 extended sources smaller than 0.5°, while recovering only 1 of 33 large diffuse sources above 0.8°, which the filter suppresses as low-frequency structure. The seeds are classified by blob scale as point-like or extended and passed to the usual HAWC likelihood fitter for final localization. The image-processing stage took under 26 CPU-hours total against more than two months for the conventional blind search, a 60× speedup (the abstract quotes up to 300×), corresponding to the reported 98.33% performance gain.","pith_inferences":["A decisive next test would freeze the 0.20° smearing radius and 5σ threshold on one set of simulations and measure recovery on independent simulations or real HAWC maps, since the paper reports performance on the same simulations used to demonstrate those parameters.","The blob scale returned by the finder could feed extended-source templates into the likelihood fit, giving the fitter a size prior that might recover some of the large diffuse sources the current filter suppresses.","At the reported speed, a natural deployment is a real-time or alert-driven scan that seeds candidate sources in freshly made significance maps without waiting for a full likelihood model.","Tuning the smearing radius to the local point-spread function of each detector would likely be needed when moving the pipeline to other instruments."],"forward_implications":["Source seeding for a HAWC region can drop from days to hours, leaving likelihood computation for refining a short candidate list.","Fainter point sources that the iterative fit misses should appear as seeds and can be added to the model manually or automatically.","Because the method works on significance images rather than detector-specific likelihoods, it can be ported to Fermi-LAT maps and to future LHAASO and SWGO data.","Large diffuse sources will need a separate search, since the Difference-of-Gaussians filter removes them along with the background.","The saved computation makes wide-area or repeated scans of the very-high-energy sky practical."],"supporting_citations":[{"why":"Supplies the maximum-likelihood framework that evaluates the seeded source candidates.","marker":"[1]"},{"why":"Provides the HAWC-accelerated likelihood plugin used to localize the seeds.","marker":"[2]"},{"why":"Describes the existing HAWC multi-source search pipeline that the new approach is compared against for speed.","marker":"[4]"},{"why":"Supplies the image-processing filters and blob-finder implementation used in the pipeline.","marker":"[5]"},{"why":"Provides the theoretical basis for the Difference-of-Gaussians filtering step.","marker":"[6]"},{"why":"Supplies the multi-scale local-extrema blob detection method the seed finder is based on.","marker":"[7]"}],"fun_headline_variants":["Image stack finds HAWC sources 60x faster than blind search","Pre-filter seeds recover 65 of 66 HAWC point sources","Difference-of-Gaussians speeds HAWC source search 60x","Image processing seeds gamma-ray sources before fit","HAWC source search cut from months to hours"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The measured speed and recovery rates assume that the fixed 0.20° smearing radius, the 5σ detection threshold, and the other cut values will behave the same way outside the 60 simulated maps on which they are demonstrated.","fun_headline_variants_meta":{"raw":{"variants":["Image stack finds HAWC sources 60x faster than blind search","Pre-filter seeds recover 65 of 66 HAWC point sources","Difference-of-Gaussians speeds HAWC source search 60x","Image processing seeds gamma-ray sources before fit","HAWC source search cut from months to hours"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000685,"raw_usage":{"total_tokens":3070,"prompt_tokens":874,"completion_tokens":2196,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":490,"completion_tokens_details":{"reasoning_tokens":2114}},"tokens_in":490,"tokens_out":2196,"duration_ms":18232,"temperature":1.0,"reasoning_tokens":2114,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T17:33:38.807493+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the pipeline with those fixed parameters on a held-out set of simulations, or on a real HAWC significance map whose sources are already known, and count how many injected or catalogued point and compact sources appear in the seed list; if recovery falls well below 65 of 66 point sources and 24 of 25 compact extended sources, or if the wall-clock advantage over the likelihood search disappears in a head-to-head test, the central claim would not transfer.","supporting_citations":[{"cited_title":"TeV Analysis of a Source Rich Region with HAWC Observatory: Is HESS J1809-193 a Potential Hadronic PeVatron?","cited_arxiv_id":"2407.08849","evidence_quote":"Supplies the image-processing filters and blob-finder implementation used in the pipeline."}],"review_version":1}