{"id":"bf445db1-ab60-44f7-aba0-313acad69fae","arxiv_id":"2501.16530","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A catalogue of 6,956 H-alpha-excess point-source candidates from S-PLUS DR4, with machine-learning grouping and proposed colour-colour classification cuts.","lead":"This paper uses the 12-filter S-PLUS southern survey to build a catalogue of 6,956 candidate point sources with an H-alpha excess, then applies clustering and random-forest classifiers to sort them into Galactic, extragalactic, and variable-star groups. The practical interest is a target list and simple colour cuts that could guide spectroscopic follow-up of these sources.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"GDS 'upper-locus' fit is the load-bearing step: if it is not the unreddened main sequence, the 5σ criterion (Eq. 1) shifts, so the 6,956-candidate count and the classification inherit unknown contamination and incompleteness.","rationale":"Identified the single load-bearing concern as the GDS stellar-locus fit because the catalogue is the central deliverable and the fit defines the selection threshold. The reader's weakest_assumption correctly points at this and I agree. A secondary concern is the circular RF/HDBSCAN validation and potential magnitude-driven UMAP clusters, but those affect the classification layer only; even if the clusters were spurious, the 6,956-candidate list would stand or fall on Eq. 1. Thus the stress-test does not change the reader's CONDITIONAL verdict: the catalogue and fit validation are required.","tokens_in":32465,"tokens_out":8979,"duration_ms":91941,"concrete_test":"A direct test is to use VPHAS+ as an independent Hα-selection in the overlapping GDS fields: run the same 5σ locus method on VPHAS+ r−NB659 photometry for the matched stars, then compute (a) the fraction of the 3,319 S-PLUS GDS candidates that are also VPHAS+ Hα-excess candidates and (b) the fraction of VPHAS+ Hα-excess candidates recovered by S-PLUS, separated by r-magnitude bin and (r−i) colour. If recovery on the reddened locus is well below ~70% while the blue part recovers well, or if a large fraction of S-PLUS candidates show no VPHAS+ excess, the 'upper-locus' assumption fails and the GDS count is biased.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central catalogue is produced by Eq. 1 in Section 3.3, where the reference is a per-field, per-magnitude-bin straight line fitted iteratively to the stellar locus. For the GDS, when two loci are present, the paper moves the fit upward to the 'upper locus' and asserts it 'generally corresponds to the unreddened main sequence', but it never overlays the Pickles tracks on a GDS field to verify this, and the procedure is not fully specified (points above the initial fit are re-fitted, so genuine Hα emitters can drag the line upward). If the upper locus is actually a foreground unreddened giant/blue population, then all thresholds shift; if it is a reddened main sequence, objects on the lower (reddened) locus require an excess equal to the two-locus separation plus 5σ before selection, systematically deleting reddened disk emitters. The only external validation (Section 4.4) reports a median r−Hα offset of 0.02 mag but MAD=0.27 mag for 793 GDS matches, which does not demonstrate that the S-PLUS candidates coincide with VPHAS+ Hα emitters; no overlap/completeness test is presented. Because the headline number (3,319 GDS candidates) depends directly on this fit, the fit's validity is the most load-bearing assumption in the paper.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents a search for Hα-excess point sources in S-PLUS DR4 using the (r−J0660) versus (r−i) colour-colour method of Witham et al. (2008). The authors first apply iterative linear fits to the stellar locus in each field and magnitude bin for the Main Survey (MS) and the Galactic Disk Survey (GDS), selecting sources that lie at least 5σ above the fit. They report 6,956 candidates, 3,637 from the MS and 3,319 from the GDS, and cross-match them with SIMBAD, SDSS, LAMOST, and VPHAS+. They then use UMAP plus HDBSCAN to cluster the MS candidates, first with 66 S-PLUS colours and then with 77 colours including WISE W1/W2, and train a Random Forest on the cluster labels to identify important colours. The paper concludes that the S-PLUS 12-filter system can separate RR Lyrae stars and eclipsing binaries from genuine Hα emitters, that optical plus infrared data improve the separation of Galactic from extragalactic sources, and that the proposed colour-colour diagrams provide tentative pre-classification criteria.","tokens_in":32738,"tokens_out":6718,"duration_ms":67969,"significance":"If the candidate catalogue is sound, this work would provide a new southern-sky list of Hα-excess point sources and demonstrate the utility of S-PLUS narrow-band J0660 photometry for finding emission-line objects and variables. The paper has several genuine strengths: the selection procedure is described in enough detail to be followed; the SDSS and LAMOST spectra in Figs. 7–10 convincingly show that at least some selected objects are genuine emitters, CVs, and QSOs with redshifted lines; and the authors are explicit about contamination by RR Lyrae and eclipsing binaries and about caution for r<13.5 mag. However, the central catalogue is not actually included or linked, the GDS selection depends on an unverified choice of the 'upper locus' as the unreddened main sequence, and the Random Forest validation is circular with respect to physical classification because it is trained on HDBSCAN labels generated from the same photometric features. These issues currently prevent the reader from independently checking the headline numbers or the physical interpretation of the clusters.","major_comments":[{"comment":"The GDS upper-locus fit is the load-bearing step for the 3,319 GDS candidates. The text states that the upper locus 'generally corresponds to the unreddened main sequence' but never demonstrates this by overlaying the Pickles main-sequence and giant tracks shown in Figs. 2–3 on a GDS field. Because the iterative procedure re-fits points above the initial line, genuine Hα emitters can drag the fitted line upward, and if the upper locus is in fact a reddened or giant-dominated population, Eq. (1) thresholds shift for every GDS source. The procedure is also not fully specified: the clipping threshold is described only as 'several σ' and the iteration count as '2 to 3'. Please overlay the Pickles tracks on representative GDS fields, quantify the offset between the initial and final fits per field and magnitude bin, specify the clipping algorithm precisely, and report how the 3,319-candidate count changes if the lower (reddened) locus is used as the reference.","section":"3.3 (Eq. 1, Fig. 5)"},{"comment":"The central catalogue is not provided. The headline result of the paper is a list of 6,956 Hα-excess candidates, but the manuscript contains no machine-readable table, no VizieR/CDS link, and no appendix table with candidate positions, photometry, errors, and flags. Without the catalogue, readers cannot verify the reported counts, the SIMBAD cross-match statistics, or the clustering metadata. Please add a full catalogue, ideally at CDS/VizieR, and describe its columns, duplicate handling, and provenance.","section":"4 and 6"},{"comment":"The Random Forest validation is circular with respect to physical classification. The model is trained on HDBSCAN labels generated from the same 66–77 photometric colours used as input features, so the reported F1 macro average of 0.95 and the feature importances quantify how well the Random Forest reproduces the cluster labels, not whether the clusters correspond to physically distinct source classes. This circularity also affects the interpretation of Groups 0–4 in Table 3 and the tentative colour cuts in Fig. 19. Please either reframe Section 5.3 as an analysis of cluster-label reproducibility or validate the clusters against external spectroscopic classifications, reporting per-class precision and recall for confirmed CVs, QSOs, RR Lyrae stars, and eclipsing binaries.","section":"5.3 (Fig. 18, Fig. 19, Table 3)"},{"comment":"The spectroscopic confirmation rates are not defined on a well-posed sample. The paper states that 'approximately 60%' of the 212 SDSS/LAMOST matches show emission lines and 'around 30%' show Hα absorption, but the denominator, the selection function, and any spectral signal-to-noise or line-strength thresholds are not given. The SDSS and LAMOST overlaps are not a random subsample of the candidate list, so these percentages cannot be interpreted as purity estimates for the full catalogue. Please provide a per-survey table of spectral classifications (emission, absorption, no line, unknown) and, if possible, report confirmation rates for a randomly selected or completeness-weighted subsample.","section":"4.3"},{"comment":"The VPHAS+ comparison does not validate the candidate selection. The 793 matched objects are used to compare r−i and r−Hα colour differences between S-PLUS and VPHAS+, yielding medians of −0.21 and +0.02 with MADs of 0.07 and 0.27, but this does not test whether S-PLUS Hα-excess candidates coincide with VPHAS+ Hα-excess sources. A meaningful external check would cross-match the S-PLUS GDS candidate list with VPHAS+ photometry and compare the Hα-excess flags, reporting the fraction of S-PLUS candidates that are also VPHAS+ emitters and vice versa.","section":"4.4 (Fig. 11)"}],"minor_comments":[{"comment":"The manuscript header contains placeholder dates ('Received September 15, 1996; accepted March 16, 1997') that should be corrected before submission.","section":"Header"},{"comment":"The filter is repeatedly written as 'J0600' in Section 4 and in the caption of Fig. 6; it should be 'J0660' consistently throughout the paper.","section":"Section 4 and Fig. 6"},{"comment":"The colour-colour diagrams in Fig. 19 are difficult to use because the pairplot panels lack clear axis labels and the tentative colour cuts are not legible in the printed figure; please replot with labelled axes and a legend for the cuts.","section":"Fig. 19"},{"comment":"The sentence 'We ended up with the optimal value of 2 and 50, respectively' does not specify which HDBSCAN parameter is 2 and which is 50; the order for min_samples and min_cluster_size should be stated explicitly.","section":"Section 5.2.1"},{"comment":"The text 'typically exhibit amplitudes higher than 0.3 to 2 magnitudes for RR Lyrae stars Chandra X-ray Observatory' contains what appears to be a citation placeholder; please replace it with proper references for RR Lyrae periods and amplitudes.","section":"Section 5.2.2"},{"comment":"There are several grammar and typographical slips, such as 'we primary goal of identify objects' in Section 2 and 'dmax,pro j' in Section 4; these should be corrected in a copy-editing pass.","section":"Section 2"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the scope of A&A and the S-PLUS data are valuable, but the gap between the catalogue claims and the material provided is the main editorial concern: the candidate list is missing, and the GDS selection rests on an unverified locus choice. The machine-learning section should be reframed as cluster analysis rather than physical classification unless external validation is added. I believe these issues are addressable in a major revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this paper does exactly what it advertises — a southern-sky catalogue of 6,956 J0660-excess point sources from S-PLUS DR4, plus photometric pre-classification diagrams. It is not a methodological breakthrough; the selection recipe (Witham et al. 2008) and the ML stack (UMAP/HDBSCAN/RandomForest) are off the shelf. The value is in the new catalogue and the S-PLUS application.\n\nWhat is new and good: the PSF photometry pipeline for the Galactic Disk Survey is a concrete addition (SExtractor/PSFex on median-subtracted images), and the spectral examples — CVs, QSOs at z 1.4–3.3, Seyferts, RR Lyrae in absorption — convincingly show that the colour selection finds real emission-line and absorption-line objects. The SIMBAD cross-match gives a plausible object mix, and the RR Lyrae explanation (differential absorption in the broad r versus narrow J0660) is physically sensible. The selection procedure is documented in enough detail to reproduce.\n\nSoft spots, in order. (1) The GDS upper-locus fit is load-bearing and under-verified. The paper asserts the upper locus 'generally corresponds to the unreddened main sequence' but never overlays the Pickles tracks on a GDS field to test it, and the iterative procedure — refitting points above the initial line — can pull the fit upward toward genuine Hα emitters. The VPHAS+ comparison (median 0.02, MAD 0.27 on r−Hα) is not an overlap or completeness test of the candidate selection; it demonstrates colour consistency only. This directly affects the 3,319 GDS candidates. (2) The Random Forest validation is circular: trained on HDBSCAN labels, its F1 = 0.95 shows the model reproduces the clusters. The SIMBAD-matched group compositions provide external anchoring (Group 1 dominated by RR Lyrae, Group 2 by eclipsing binaries) but not a quantitative purity estimate. (3) The machine-readable catalogue is missing from the preprint — a central omission for a catalogue paper. (4) Minor: the tentative colour cuts in Fig. 19 are drawn but not tabulated, and the paper itself warns about r < 13.5 while the GDS magnitude distribution peaks exactly there.\n\nThe core claim — that S-PLUS DR4 photometry yields thousands of genuine Hα-excess candidates — holds up. The selection is standard and the spectral examples confirm it. The quantitative classification claims and the exact candidate count need the catalogue release and an independent spectroscopic validation of the ML groups.\n\nVerdict: send it to review. The paper deserves a serious referee; the fixes — release the catalogue, justify the GDS locus, or temper the claims — are within the authors' reach.","headline":"A useful southern-sky candidate catalogue, with a load-bearing GDS locus assumption and circular ML validation that a revision should fix.","tokens_in":33445,"tokens_out":4141,"would_cite":true,"duration_ms":39061,"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":"S-PLUS data yield 6,956 candidate H-alpha-excess point sources in the southern sky.","keywords":["H-alpha excess","S-PLUS survey","J0660 narrowband filter","colour-colour selection","UMAP","HDBSCAN","RR Lyrae","QSO and AGN classification"],"falsifier":"Randomly select a few hundred candidates from each survey and take optical spectra, counting the fraction with an emission or absorption feature inside J0660; if that fraction falls far below the paper's quoted roughly 60% emission plus 30% absorption among spectral matches, the 5-sigma locus criterion is not selecting what it claims. Re-running the fit with the threshold varied from 3 to 7 sigma should also change the candidate counts smoothly, not in a cliff that would signal an unstable fitted locus.","tokens_in":32285,"feed_emoji":"🔭","tokens_out":8223,"duration_ms":76373,"temperature":0.7,"pith_summary":"The paper claims that the S-PLUS survey's narrowband J0660 filter, together with the broadband r and i filters, can reveal point sources whose light is modified by hydrogen-alpha emission or by other emission lines shifted into that filter. Applying a per-field stellar-locus colour excess method to the survey's fourth data release, it builds a catalogue of 6,956 candidates: 3,637 from the high-latitude Main Survey and 3,319 from the Galactic Disk Survey. Feeding all 12 S-PLUS colours to unsupervised clustering separates Galactic from extragalactic sources and isolates RR Lyrae stars, whose apparent excess arises from H-alpha absorption combined with the survey's sequential filter exposures. Adding mid-infrared colours sharpens the separation so cataclysmic variables can be told apart from quasars at redshift near 1.35, and a feature-ranking model distills the discrimination into simple colour-colour diagrams with tentative cuts. If correct, the work gives the southern sky a homogeneous H-alpha-excess candidate list and cheap pre-classification criteria for follow-up spectroscopy.","feed_headline":"S-PLUS flags 6,956 H-alpha-excess sources in southern sky","feed_subtitle":"Twelve-filter photometry plus machine learning separates true emitters from RR Lyrae and redshifted quasars.","key_machinery":"The central machinery is the (r−J0660) versus (r−i) colour-colour excess criterion: a linear fit to the stellar locus per field and magnitude bin, refined by iterative sigma clipping, defines the expected colour of a source without H-alpha; the selection condition (r−J0660)obs − (r−J0660)fit ≥ 5·σest flags candidates, where σest combines the fit scatter with the photometric errors of both colours. For the Galactic Disk Survey the iterative fit is moved upward toward the unreddened main sequence to handle differential reddening. Classification uses UMAP dimensionality reduction on 66 S-PLUS colour indices followed by HDBSCAN clustering, with a second pass adding two infrared bands to make 77 colours; a Random Forest trained on the cluster labels ranks the colours and provides the basis for simplified colour-colour cuts.","core_discovery":"The central claim is that H-alpha-excess point sources over the southern sky can be identified from S-PLUS DR4 photometry and grouped into astrophysical classes using colours alone. In the (r−J0660) versus (r−i) plane, non-emitting stars define a locus; the authors fit a line to that locus in each field and magnitude bin with iterative sigma clipping, then flag any source lying at least five estimated standard deviations above the fit. In crowded Galactic-disk fields with two loci produced by differential reddening, the fit is iterated upward to the presumed unreddened main sequence. Cross-matches show the selected objects are emission-line stars, young stellar objects, binaries, cataclysmic variables, planetary nebulae, RR Lyrae stars, quasars, and active galactic nuclei; for some quasars the J0660 excess is actually Mg II, C III], C IV, [O III], or H-beta redshifted into the filter. Clustering on 66 S-PLUS colours separates a quasar/AGN-dominated group from a group of RR Lyrae stars and binaries, and adding infrared colours resolves the cataclysmic-variable/quasar confusion. The authors deliberately call the sources 'H-alpha-excess' rather than 'H-alpha emitters,' since absorption lines and variability can produce the same photometric signature.","pith_inferences":["The same per-field locus-fitting procedure should transfer to the northern-sky twin survey with the same filter set; a direct test is cross-matching the colour criteria in overlap regions.","Because the 5-sigma threshold is fixed and estimated per field, catalogue purity and completeness are likely patchy; a continuous probabilistic excess score would give each candidate an individual confidence.","The clean separation of RR Lyrae stars suggests the 12-filter sequence itself encodes short-period variability information, which could be mined for variability beyond the H-alpha-excess sample.","The remaining optical-only confusion between cataclysmic variables and quasars at z ~ 1.35 marks a boundary for photometric classification that infrared colours only partly resolve; adding other line-sensitive bands could push it further."],"forward_implications":["The 6,956-entry catalogue gives southern-sky surveys a compact point-source list for spectroscopic follow-up of H-alpha-excess candidates.","Any S-PLUS field with clean photometry can be processed with the same per-field locus fit, so the method extends to future S-PLUS data releases.","RR Lyrae stars and eclipsing binaries will keep appearing in sequential narrowband H-alpha selections because phase sampling and H-alpha absorption mimic excess; the paper shows how to recognise that population.","Quasars and AGN at particular redshifts masquerade as H-alpha emitters when Mg II, C III], C IV, [O III], or H-beta lands in J0660, so infrared colours are needed to separate extragalactic from Galactic candidates.","The simplified colour-colour diagrams distilled from the clustering allow pre-classification with a few filters, reducing the need for complete 12-band photometry."],"supporting_citations":[{"why":"Supplies the iterative stellar-locus fitting method and the C-sigma excess criterion that the S-PLUS selection adapts to (r-J0660) versus (r-i).","marker":"Witham et al. (2008)"},{"why":"Provides S-PLUS DR4 images, catalogues, and flux calibration, including the synthetic-photometry zero points the selection depends on.","marker":"Herpich et al. (2024)"},{"why":"Defines the S-PLUS 12-filter system and the J0660 narrowband filter whose transmission sets which lines create the excess.","marker":"Mendes de Oliveira et al. (2019)"},{"why":"Supplies the PStotal photometry and the magnitude, error, and morphological limits used to build the Main Survey point-source sample.","marker":"Almeida-Fernandes et al. (2022)"},{"why":"Provides the synthetic stellar spectra convolved with S-PLUS filters to draw the main-sequence and giant tracks that anchor the colour-colour diagrams.","marker":"Pickles (1998)"},{"why":"Supplies the UMAP algorithm used for dimensionality reduction of the 66 and 77 colour features.","marker":"McInnes et al. (2020)"},{"why":"Supplies the HDBSCAN clustering algorithm that defines the source groups.","marker":"Campello et al. (2013)"},{"why":"Supplies the infrared catalogue and the W1/W2 bands whose addition sharpens the Galactic/extragalactic separation.","marker":"Wright et al. (2010)"},{"why":"Supplies the Random Forest classifier used to rank colour importance and derive the tentative colour cuts.","marker":"Breiman (2001)"}],"fun_headline_variants":["S-PLUS finds 6,956 H-alpha-excess sources via ML","6,956 H-alpha-excess sources classified from S-PLUS","Machine learning maps 7,000 H-alpha-excess candidates","Southern survey flags 6,956 H-alpha-excess point sources","Photometric survey + ML separates H-alpha-excess classes"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The candidate list rests on the fitted lines in the (r−J0660) versus (r−i) plane tracing the colours of genuinely non-emitting stars, and, in the Galactic Disk Survey, on the upper locus being the unreddened main sequence; if reddening or calibration shifts those fits, the 5-sigma criterion adds contaminants or drops real emitters.","fun_headline_variants_meta":{"raw":{"variants":["S-PLUS finds 6,956 H-alpha-excess sources via ML","6,956 H-alpha-excess sources classified from S-PLUS","Machine learning maps 7,000 H-alpha-excess candidates","Southern survey flags 6,956 H-alpha-excess point sources","Photometric survey + ML separates H-alpha-excess classes"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000575,"raw_usage":{"total_tokens":2827,"prompt_tokens":1170,"completion_tokens":1657,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":786,"completion_tokens_details":{"reasoning_tokens":1565}},"tokens_in":786,"tokens_out":1657,"duration_ms":11637,"temperature":1.0,"reasoning_tokens":1565,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T12:33:47.936239+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Randomly select a few hundred candidates from each survey and take optical spectra, counting the fraction with an emission or absorption feature inside J0660; if that fraction falls far below the paper's quoted roughly 60% emission plus 30% absorption among spectral matches, the 5-sigma locus criterion is not selecting what it claims. Re-running the fit with the threshold varied from 3 to 7 sigma should also change the candidate counts smoothly, not in a cliff that would signal an unstable fitted locus.","supporting_citations":[{"cited_title":"R., Knigge , C., Drew , J","cited_arxiv_id":null,"evidence_quote":"Supplies the iterative stellar-locus fitting method and the C-sigma excess criterion that the S-PLUS selection adapts to (r-J0660) versus (r-i)."},{"cited_title":"R., Almeida-Fernandes , F., Oliveira Schwarz , G","cited_arxiv_id":null,"evidence_quote":"Provides S-PLUS DR4 images, catalogues, and flux calibration, including the synthetic-photometry zero points the selection depends on."},{"cited_title":"2020, UMAP: Uniform Manifold Approximation and Projection for Dimension Reduction","cited_arxiv_id":null,"evidence_quote":"Supplies the UMAP algorithm used for dimensionality reduction of the 66 and 77 colour features."}],"review_version":1}