{"id":"983df9ed-e141-4bcf-ab07-46162dd6fdf5","arxiv_id":"1909.01929","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":3,"one_line_summary":"A preliminary image-processing and machine-learning pipeline classifies cosmic-ray tracks in CMOS/CCD camera images into muon-like, electron-like, and noise, and estimates muon track azimuth.","lead":"This paper describes a method for spotting and classifying cosmic-ray hits in ordinary camera photos, using the shape of the trace left on the sensor and camera metadata. It is a concept and early implementation for a global network of smartphones and webcams acting as cheap cosmic-ray detectors.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No calibration or ground truth connects image-shape classes to particle species or track orientation to physical azimuth; the paper's own §3.6 says confirmation requires radiation-source or simulation studies.","rationale":"The reader's CONDITIONAL verdict already identifies the same load-bearing weakness: the absence of any calibration connecting image morphology to particle identity and orientation angle to physical azimuth. My stress-test pass finds no additional fatal flaw beyond this, and no reason to move the verdict to ACCEPT or REJECT. The paper is explicitly a preliminary concept note, it links public code, and it contains internal caveats in §3.6 that acknowledge the need for radiation-source, simulation, or multi-device statistical confirmation. The abstract overstates the azimuth capability, but this is best handled by requiring validation rather than by rejecting the contribution outright. The proposed test — injecting simulated muon tracks with known direction into real dark frames and measuring classification and azimuth errors — would directly settle whether the shape-to-species and orientation-to-azimuth mappings hold. Until such a test is run, the central claim remains unverified but not internally inconsistent, so CONDITIONAL is the appropriate verdict and no adjustment is needed.","tokens_in":6309,"tokens_out":2799,"duration_ms":32589,"concrete_test":"Take a set of ESO dark frames and inject simulated muon tracks with known 3D momentum directions (e.g., from CORSIKA or GEANT4 with a sensor model), run the complete pipeline from §3.2–§3.4, and report the classification confusion matrix and the azimuth residual between the estimated and injected track azimuth. If the azimuth RMS error is comparable to or larger than the ~30° compass uncertainty, or if track-like/electron-like classification accuracy is not substantially better than the shape-threshold baseline, then the central claim fails. An accelerator or radioactive-source beam test with known particle type and direction would be an even more direct check.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim in the abstract — azimuth estimation for muon-like events — depends on two mappings that the paper asserts but never validates: (i) §3.4 states 'It suppose that the track-like hits is caused by muons hits in angle to sensor plane' and worm-like by beta radiation, without calibration against known particle sources, simulations, or a labeled dataset; and (ii) orientation of the image trace is taken to give the muon azimuth, but §3.6 notes that correct azimuth requires metadata and that compass accuracy is often ~30°, and 'More experiments with physical radiation source or cosmic-ray peak simulations or statistical analysis from more devices are required in order to confirm it.' The histograms in Fig. 4 are qualitative, and the authors themselves identify track-like artifacts at exactly 0°, 45°, and 90° that could contaminate the orientation distribution. No confusion matrix, efficiency, false-positive rate, or azimuth residual is reported, and the CNN classifier of [18] that could serve as a baseline is only cited. Thus the capability claim 'estimates azimuth' is unsupported by the evidence presented, even though the paper is honestly framed elsewhere as a preliminary concept.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents a concept and preliminary implementation for recognizing cosmic-ray events in CMOS/CCD images, using data from the CREDO project and ESO dark frames. The proposed pipeline detects bright hits, crops them, extracts geometric features (ellipticity, solidity, orientation, length, width) and metadata, and classifies events as muon-like, electron-like, or other/noise using hand-set morphology rules. For muon-like events, the abstract also claims the method estimates the azimuth of the muon track. The paper describes the processing chain, refers to publicly available code, and illustrates the output with histograms from a single CREDO device. The authors explicitly frame the work as a proof-of-concept that needs further validation, particularly in the statistical-analysis section.","tokens_in":6471,"tokens_out":3025,"duration_ms":32507,"significance":"If the central claims could be validated, this would offer an appealing capability: a distributed network of ordinary cameras acting as cosmic-ray muon detectors with directional information. The paper has the virtue of reproducibility: references [15] and [16] provide working code for feature extraction and analysis, and the data sources (CREDO, ESO) are public. However, the capability claims are currently supported only by morphological assumptions, not by quantitative evidence. No classification accuracy, ground truth, confusion matrix, detection efficiency, or azimuth residual is reported, and the paper itself acknowledges that confirmation with radiation sources or simulations is required. The contribution is therefore best seen as an openly documented pipeline whose central claims require demonstration.","major_comments":[{"comment":"The classification into muon-like, electron-like, and other events rests on untested physical assumptions. The text states 'It suppose that the tack-like hits is caused by muons hits in angle to sensor plane, worm-like is caused by β-radiation,' but no calibration against known particle sources, simulations, or a labeled dataset is provided. The paper reports no confusion matrix, classification efficiency, or false-positive rate. Because the abstract's central claim about identifying muon-like events depends entirely on this morphology-to-particle mapping, the classification step needs at least a basic validation or an explicit rephrasing as an unverified hypothesis.","section":"§3.4 and Abstract"},{"comment":"The claim that the method 'estimates azimuth of a muon track' is not supported by the evidence presented. The bottom-left histogram in Fig. 4 plots track orientations, but there is no comparison to independently known azimuths, no compass/GPS validation, and no azimuth residual or angular-error analysis. The authors themselves note that compass accuracy is often about 30°, that artifacts appear at exactly 0°, 45°, and 90°, and that 'more experiments with physical radiation source or cosmic-ray peak simulations or statistical analysis from more devices are required in order to confirm it.' With the currently shown data, the trace orientation is measured, but the physical azimuth is not demonstrated to be estimated.","section":"§3.6 and Abstract"},{"comment":"The statistical analysis is qualitative: the histograms illustrate distributions but do not quantify separation power, selection thresholds, or event counts. In particular, the ellipticity and solidity values that define spot-like, track-like, and worm-like events are given only as illustrative examples in Fig. 3c, not as explicit threshold values or as a reproducible decision rule. This makes it impossible for a reader to reproduce the classification or assess its sensitivity and efficiency.","section":"Fig. 4"}],"minor_comments":[{"comment":"There is a typo: 'tack-like' should be 'track-like'. Also, in §3.6 'mouns' should be 'muons'.","section":"§3.4"},{"comment":"Reference [10] is listed as 'The scikit-image project home page', but the URL points to http://photutils.readthedocs.io/; please make the citation consistent with the actual library used for feature extraction.","section":"References"},{"comment":"The phrase '2 millions images' and the compound 'cameralike' in the abstract are awkward; please rephrase as '2 million images' and 'camera types'.","section":"Abstract and §4"},{"comment":"Each panel of Fig. 4 should state the number of events plotted, the device model, and the threshold criteria used to select the subsamples, so that the histograms can be interpreted and reproduced.","section":"Fig. 4"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThis is an honest preliminary methods note from the CREDO collaboration, not a validated result. The spot/track/worm classifier is not new—it comes from earlier DECO/CRAYFIS work and Meehan et al.'s CNN, which the authors cite. The genuinely new piece is applying this to CREDO's heterogeneous consumer-camera database and adding a heuristic azimuth estimate for track-like events. That new piece is exactly where the evidence stops.\n\nThe paper does well on transparency: the pipeline is described clearly, the code is linked, and the authors openly flag artifacts (0/45/90 degree tracks, hot pixels, badly covered cameras). The figures give a fair sense of the data quality problems. For a conference proceedings describing work in progress, this level of disclosure is fine.\n\nThe soft spot is the azimuth claim. The abstract says the method estimates azimuth for muon-like events, but nothing in the paper validates that mapping. Track orientation can only give azimuth if two assumptions hold: shape class implies particle species (track-like = muon, worm-like = beta), and image orientation plus compass/GPS gives absolute azimuth. The authors themselves say in §3.6 that compass accuracy is often ~30 degrees and that confirmation requires radiation-source or simulation studies. There is no confusion matrix, no ground-truth tracks, no comparison against the CNN baseline they cite. So the central capability claim is unsupported. This is not a fatal error—it is an unvalidated proof of concept—but the abstract slightly oversells it.\n\nI did not find fabrication. The code links and data descriptions are concrete, and the classification literature is cited properly. The circularity concern is real but mild: the classes are defined by the same shape features used to interpret them, so the particle assignment rests on an assumption rather than on external calibration.\n\nWho is this for? Someone working on distributed cosmic-ray detection with commodity cameras, or on CREDO's data pipeline, will find a useful pipeline description and a clear list of open problems. It is not yet citable evidence that commodity cameras can reconstruct muon azimuths.\n\nRecommendation: I would send it to peer review—the question is worth asking and the paper is fixable—but the review must insist on validation: synthetic tracks, irradiation with a known source, or comparison with DECO's tagged data. Without that, the azimuth claim should be explicitly downgraded to a proposal.","headline":"A transparent preliminary methods note from CREDO; the azimuth-estimation claim is the weak spot and needs validation before it becomes a capability.","tokens_in":7134,"tokens_out":2719,"would_cite":false,"duration_ms":26166,"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":"Ordinary camera images carry enough shape information to classify cosmic-ray hits as muon-like, electron-like, or noise, and the muon-like track orientation estimates azimuth.","keywords":["cosmic-ray detection","CMOS sensors","CCD sensors","muon tracks","image morphology","ellipticity","solidity","particle classification"],"falsifier":"A controlled calibration: expose the same camera to a beam of known muons and to a beta source, and check whether muons at shallow angles always give track-like traces (ellipticity above 0.6, solidity above 0.7) and betas always give worm-like traces; any overlap in morphology between the two sources would break the classification. For the azimuth claim, compare track orientation angles against a coincidence telescope co-mounted with the camera; the spread of the residuals measures how much of the orientation is real muon azimuth versus noise.","tokens_in":6087,"feed_emoji":"📷","tokens_out":8629,"duration_ms":75501,"temperature":0.7,"pith_summary":"This paper argues that a single bright trace left in a digital photo can identify a cosmic-ray particle and roughly indicate where it came from. The authors propose that shape features of a hit, especially ellipticity and solidity, separate muon-like tracks from beta-like worm traces and from noise, and that the orientation angle of a muon-like track estimates the muon's azimuth. As proof of concept they process telescope dark frames and a citizen-contributed set of over two million camera images. If the method holds, ordinary smartphones, webcams, and IoT cameras become a worldwide, zero-hardware cosmic-ray observatory.","feed_headline":"Camera traces reveal cosmic-ray muons and their azimuth","feed_subtitle":"Shape features like ellipticity and solidity classify sensor hits; track angle estimates the muon's arrival direction.","key_machinery":"The load-bearing object is the per-hit feature vector: ellipticity, marked area, convex area, orientation, track length, and track width, derived from the cropped image and combined with acquisition metadata. Solidity, defined as marked area over convex area, is the companion of ellipticity: the pair separates clean straight tracks from bent worm-like traces and blobs. The same orientation angle used to describe the ellipse bounding the hit becomes, once a hit is classified as muon-like, the estimate of the muon's azimuth. The paper also exploits the fact that hot pixels stay at fixed coordinates from frame to frame, providing a cheap way to cut them out before classification.","core_discovery":"The paper claims that a cosmic-ray hit on a CMOS/CCD sensor can be recognized and typed from the geometry of the trace it leaves. Spot-like hits (ellipticity near 1 and solidity near 1), track-like hits (ellipticity above 0.6 and solidity above 0.7) attributed to muons crossing the sensor plane at an angle, and worm-like hits (ellipticity around 0.5 and solidity below 0.7) attributed to beta radiation are the three morphological kinds. A machine-learning classifier fed with these shape features plus metadata such as camera model, pixel size, RAW versus down-sampled format, GPS location, timestamp, and exposure assigns each cropped hit to muon-like, electron-like, or other; for muon-like events the orientation angle of the track is read as the muon's azimuth. The authors present this as a concept and preliminary implementation, tested on telescope dark frames and on a citizen-contributed set of more than two million camera images.","pith_inferences":["A testable extension the paper does not run: calibrate the three morphological classes against a known source, converting 'muon-like' into 'muon'.","Since the paper notes that transmission artifacts cluster at exactly 0°, 45°, and 90°, an orientation histogram could double as a per-device artifact veto and as a cross-check on any azimuth anisotropy.","If the morphological mapping holds, the same features may carry energy information, so normalizing marked area and track width by pixel size could turn the distributed camera network into a crude energy-resolving detector.","The azimuth claim combines two sub-claims: track angle equals muon azimuth on the sensor, and the sensor's absolute orientation is known; with compass errors around 30°, absolute azimuth may need a post-hoc orientation fix from recognizable sky objects in the image."],"forward_implications":["If the shape-to-particle mapping holds, a camera's ordinary images become cosmic-ray science data with no new hardware, only software.","Muon azimuths estimated from single images could feed global anisotropy studies once compass and GPS metadata are sufficiently accurate.","Astronomical dark frames, routinely taken for calibration, become a free and large dataset of cosmic-ray interactions.","Per-device auto-calibration becomes essential but tractable, since each sensor's pixel size and format change the scale of the same morphological signal.","The same pipeline can act as a filter, separating true ionizing-particle hits from hot pixels and fake covered-camera artifacts."],"supporting_citations":[{"why":"Establishes that cosmic rays create recognizable noise signatures in astronomical CCD imagers, the premise for the whole detection scheme.","marker":"[1]"},{"why":"Shows muons hitting a camera sensor appear as dots or tracks depending on the angle to the sensor plane, grounding the shape-to-particle mapping.","marker":"[2]"},{"why":"Supplies the multi-million-image citizen-camera dataset that the classification pipeline is designed to analyze.","marker":"[3]"},{"why":"Supplies the telescope dark frames from which cosmic-ray hits are extracted and cropped for analysis.","marker":"[4]"},{"why":"Prior work characterizing CCDs with cosmic-ray event shapes, supporting the morphological spot and track categories used here.","marker":"[14]"},{"why":"Baseline that classified hits as spot, track, or worm with a convolutional network; this paper's ellipticity and solidity features are offered as a simpler route to the same classes.","marker":"[18]"},{"why":"Shower simulation used to justify the expected mix of particle types behind the three-way classification.","marker":"[20]"}],"fun_headline_variants":["Classifying cosmic-ray hits by trace shape on camera sensors","Machine learning sorts muons, electrons, noise in camera images","Sensor traces reveal muon tracks and their direction","Cosmic-ray events typed by geometric shapes in camera data","Camera pixels become particle detectors via shape analysis"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that trace shape maps one-to-one to particle identity—a high-ellipticity, high-solidity track means a muon, a worm-like trace means beta radiation, and a spot could be either—and that the trace's orientation angle equals the muon's azimuth, a mapping the paper asserts without calibrating it against known particle sources or simulations.","fun_headline_variants_meta":{"raw":{"variants":["Classifying cosmic-ray hits by trace shape on camera sensors","Machine learning sorts muons, electrons, noise in camera images","Sensor traces reveal muon tracks and their direction","Cosmic-ray events typed by geometric shapes in camera data","Camera pixels become particle detectors via shape analysis"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000183,"raw_usage":{"total_tokens":1323,"prompt_tokens":966,"completion_tokens":357,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":582,"completion_tokens_details":{"reasoning_tokens":280}},"tokens_in":582,"tokens_out":357,"duration_ms":4175,"temperature":1.0,"reasoning_tokens":280,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T05:04:10.320955+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A controlled calibration: expose the same camera to a beam of known muons and to a beta source, and check whether muons at shallow angles always give track-like traces (ellipticity above 0.6, solidity above 0.7) and betas always give worm-like traces; any overlap in morphology between the two sources would break the classification. For the azimuth claim, compare track orientation angles against a coincidence telescope co-mounted with the camera; the spread of the residuals measures how much of the orientation is real muon azimuth versus noise.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes that cosmic rays create recognizable noise signatures in astronomical CCD imagers, the premise for the whole detection scheme."},{"cited_title":"Measurement of camera image sensor depletion thickness with cosmic rays","cited_arxiv_id":"1511.00660","evidence_quote":"Shows muons hitting a camera sensor appear as dots or tracks depending on the angle to the sensor plane, grounding the shape-to-particle mapping."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the multi-million-image citizen-camera dataset that the classification pipeline is designed to analyze."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the telescope dark frames from which cosmic-ray hits are extracted and cropped for analysis."},{"cited_title":"Fisher-Levine and A","cited_arxiv_id":null,"evidence_quote":"Prior work characterizing CCDs with cosmic-ray event shapes, supporting the morphological spot and track categories used here."},{"cited_title":"The particle detector in your pocket: The Distributed Electronic Cosmic-ray Observatory","cited_arxiv_id":"1708.01281","evidence_quote":"Baseline that classified hits as spot, track, or worm with a convolutional network; this paper's ellipticity and solidity features are offered as a simpler route to the same classes."},{"cited_title":"Heck and T","cited_arxiv_id":null,"evidence_quote":"Shower simulation used to justify the expected mix of particle types behind the three-way classification."}],"review_version":1}