{"id":"25a6cb62-29a7-4946-8735-837141f03b8c","arxiv_id":"2501.00094","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Clustering of 4FGL-DR3 blazar properties yields 111 changing-look blazar candidates, 67 of them new, most labeled as FSRQs.","lead":"Astronomers applied a clustering algorithm to gamma-ray blazar data and found 67 new candidates for objects that may switch between two optical spectral types over time. If confirmed, the list gives observers concrete targets for studying how supermassive black holes and their jets change state.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 111-CLBC claim rests on the authors' own TCLB labels, which also anchor the parameter choice and the ARI-selected OPCs; an independent ground-truth catalog is needed to confirm the intermediate population.","rationale":"The reader's weakest assumption correctly identifies the TCLB label dependence. I agree that this is the most load-bearing issue. The paper's own text acknowledges the arbitrary ARI threshold and the NbClust disagreement for two of four OPCs, which means the exact 111 number is not stable even within the authors' framework. The central scientific claim—that a distinct intermediate population exists and that the 67 new sources are the best candidates—would be much stronger with an independent label set or a prediction that does not use CLB labels at all (e.g., fitting a 3-component GMM and checking whether the third component's membership is enriched in CLBs). The WISE check is a nice consistency test but not a validation of the candidate list. Verdict remains CONDITIONAL: the catalog is a useful follow-up target, but the population claim needs external confirmation.","tokens_in":16505,"tokens_out":5000,"duration_ms":69332,"concrete_test":"Re-run the complete pipeline with a CLB ground-truth list compiled only from published literature independent of the TCLB catalog (e.g., Foschini et al. 2021/2022, Xiao et al. 2022, Peña-Herazo et al. 2021, Mishra et al. 2021). Keep the 8 parameters and mclust settings unchanged; reselect OPCs by ARI > 0.610 and compute the intersection candidate list. If the selected OPCs or the final list change substantially (e.g., fewer than 80 of the 111 sources survive, or the OPCs no longer include CD/lambda), the headline result is an artifact of the TCLB labels. Conversely, if a comparable intermediate group and a largely overlapping candidate list emerge, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 2.1 takes the 105 'known CLB' labels from the authors' TCLB catalog (Kang et al. 2024, in preparation; Zenodo). These same labels do double duty: they guided the choice of the 8 parameters in Kang et al. 2024, and they are the ground truth for the ARI scores that select the four OPCs in Section 3 (ARI > 0.610). The mclust clustering itself is unsupervised, but every downstream choice—which 4 of 255 subsets to report, which clusters to call 'CLBC', and the final intersection rule—is made against this internal label set. If the TCLB catalog is incomplete or biased toward sources with intermediate gamma-ray and accretion properties, the OPCs and the 111-candidate list will encode that bias rather than a real population. The fragility is visible in the paper's own robustness check: Section 4 admits the ARI threshold is arbitrary; using ARI > 0.626 keeps only No.68 and No.158, and the intersection then yields 119 candidates, not 111, while the union of all four OPCs gives 217. The WISE color-color check (Figure 7) is independent in parameter space, but it uses only 74/111 candidates and does not break the label-anchored selection of which sources are candidates. Thus the central quantitative claim is not yet free of its training labels.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper uses the mclust Gaussian Mixture Modelling package to cluster 2250 blazars from 4FGL-DR3 using subsets of eight physical parameters (Gamma_ph, alpha_ph, HR34, HR45, CD, Ldisk, Ldisk/LEdd, and z). For each of the 255 non-empty parameter subsets, the authors fit a three-component EVV mixture, evaluate the agreement with their previously compiled TCLB labels via the adjusted Rand index (ARI), and select four 'optimal parameter combinations' (No.68, No.89, No.124, No.158) with ARI > 0.610. Combining the clustering results of these four subsets by intersection yields 111 changing-look blazar candidates (CLBCs): 44 previously known CLBs, 56 sources labeled FSRQ in 4FGL-DR3, and 11 sources labeled BL Lac. The authors conclude that the CLBC group is located between FSRQs and BL Lacs in the parameter space and in the WISE color-color diagram, and they provide a machine-readable table of candidates.","tokens_in":16750,"tokens_out":2784,"duration_ms":29223,"significance":"If the claimed intermediate population is real, the 67 new CLBCs, most of which are catalogued as FSRQs, would be observationally valuable targets for spectroscopic monitoring of changing-look behaviour. The paper provides a reproducible clustering pipeline and releases machine-readable tables, which is a strength. However, the central quantitative claim of 111 CLBCs is not yet robust: the ARI threshold is explicitly admitted to be arbitrary, two of the four selected subsets are preferred with two clusters rather than three by NbClust, and the intersection rule that produces 111 rather than the 217-member union is not independently justified. The paper also uses the authors' own TCLB labels both to anchor the parameter choice and as the ground truth for cluster evaluation, so the ARI does not provide fully independent validation of the intermediate population. With additional robustness analyses and external spectral follow-up, the candidate list could become a useful resource, but as it stands the headline number should be treated with caution.","major_comments":[{"comment":"The selection of the four OPCs depends on an arbitrary ARI threshold of 0.610, as acknowledged in §4. If the threshold is raised to 0.626, only subsets No.68 and No.158 survive, and the intersection of those two subsets yields 119 CLBCs, not 111 (Table 3). Because the headline count changes by 8 candidates under a small threshold shift, the paper needs to justify the threshold, for example by comparing against a null distribution of ARI values from permutation or by reporting the final catalog as a function of threshold.","section":"§3, §4, Table 3"},{"comment":"The NbClust analysis does not support three clusters for two of the four OPCs: for No.89, 8 criteria favor two clusters, and for No.124, 10 criteria favor two clusters, while only No.68 and No.158 favor three clusters. The paper's interpretation of the CLBC group as a physical intermediate population is thus based on an imposed three-component solution for half of the selected subsets. The authors should either use a cluster-number selection that is consistent across subsets or explicitly discuss why forcing three components is physically necessary despite the NbClust result.","section":"§4, Figure 6"},{"comment":"The 105 TCLB labels are used in a double role: the eight parameters were selected in Kang et al. (2024) because known CLBs occupy intermediate values, and the same TCLB labels serve as the ground truth for the ARI scores that select the four OPCs. This circularity means that the ARI-based 'validation' partly encodes the intermediate-position hypothesis rather than independently confirming it. An external validation step—for example, optical spectroscopy of the 67 new candidates or a stability analysis that does not use the TCLB labels—is needed before the existence of a distinct intermediate CLBC population can be considered established.","section":"§2.1, §3"},{"comment":"The final 111-candidate catalog is obtained by intersecting the four OPC cluster assignments, while the union of the four subsets contains 217 sources (153 new candidates). The paper does not justify why intersection is the correct combination rule, and the choice strongly affects the results. Given that the OPC selection itself is threshold-dependent, the authors should present the sensitivity of the final catalog to both the combination rule (union, majority vote, intersection) and the OPC threshold, and should report which sources are robust across all reasonable choices.","section":"§3, Table 3"}],"minor_comments":[{"comment":"There are several typographical and grammar issues, including 'greater then 0.610' in the abstract and 'M clust' in the title; the manuscript would benefit from a careful language edit.","section":"Abstract, Section 1"},{"comment":"The text refers to 'subsets of No.98' when discussing NbClust results; from Figure 6 and the surrounding context this appears to be a typo for No.89.","section":"§4"},{"comment":"The WISE color-color check uses only 74 of the 111 CLBCs, and the paper does not discuss whether the missing 37 sources could bias the claimed intermediate location in WISE color space; this limitation should be stated explicitly.","section":"§3, Figure 7"},{"comment":"The notation for the accretion parameters is inconsistent: the text uses Ldisk and Ldisk/LEdd, while Figures 3 and 4 use labels such as 'LD_Ledd' and 'log10ADL'; unifying the notation would improve readability.","section":"§2.1"},{"comment":"The definition of the number of subsets is given in words ('the number of subsets of a set of n elements excluding the empty set'); stating the formula 2^n - 1 = 255 would be clearer.","section":"§2.2"}],"recommendation":"major_revision","confidential_remarks":"The manuscript's central catalog depends heavily on the authors' own TCLB catalog, which is described as 'in preparation' and is archived on Zenodo but not yet peer-reviewed. This makes independent checking difficult and increases the need for the robustness analyses requested in the major comments. If the authors can show that the 111-candidate list is stable under threshold, cluster-number, and combination-rule choices, or can provide external spectroscopic validation for a subset of the new candidates, the paper could become suitable for publication in ApJ."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short take: the new thing here is the list of 67 new changing-look blazar candidates and the claim that they populate a distinct intermediate region between FSRQs and BL Lacs. That broad claim is plausible and survives contact with the data. The exact count of 111 is not robust, and the paper would be stronger if the authors stopped leaning on it.\n\nWhat the paper does well: it takes a simple, transparent approach (mclust on 255 subsets of their 8 parameters), reports the ARI values, and honestly flags that the ARI cut is arbitrary. The WISE color-color check for 74 of the candidates is genuinely independent and shows the CLBC group sitting between BZQ and BZB, which corroborates the main qualitative result. The authors also acknowledge the sample-size limitation and the selection effects in the CD/Ldisk measurements.\n\nThe soft spots are real but not fatal. The 8 parameters were chosen in Kang et al. 2024 because known CLBs occupy an intermediate position, and the same known-CLB labels serve as ground truth for the ARI scores that pick the four optimal subsets. So the clustering is not fully blind. The stability of the headline number is poor: with ARI > 0.626 you get 119 candidates from two subsets; the union of all four subsets gives 217. The intersection rule that yields 111 is never justified, and NbClust actually prefers 2 clusters for two of the four OPCs. The stress-test note is right that the central quantitative claim is not yet free of its training labels.\n\nStill, this is an honest paper that does what a candidate-search paper should: it gives observers concrete targets, and the intermediate-position claim is supported by an independent check. The mismatch with Zhang et al. 2022 (mostly FSRQs instead of mostly BL Lacs) is interesting and worth following up.\n\nMy recommendation: send it to peer review. The authors should be asked to release the per-source cluster assignments, justify the intersection rule, and ideally test against an independent CLB catalog. The exact number matters less than the list and the qualitative claim.","headline":"A useful candidate list, but the headline 111 count should not be quoted: the intermediate-position claim survives, the exact number does not.","tokens_in":17374,"tokens_out":2531,"would_cite":false,"duration_ms":29636,"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":"The paper claims that Gaussian mixture clustering of 4FGL-DR3 blazars recovers a distinct intermediate population of changing-look blazar candidates, yielding 111 candidates, 67 of them new.","keywords":["changing-look blazars","Gaussian mixture model","clustering","4FGL-DR3","blazar classification","FSRQ","BL Lac objects","active galactic nuclei"],"falsifier":"Take the 67 new candidate sources and obtain optical spectra at two or more epochs; if most show no change in broad emission-line equivalent width across the 5 Å FSRQ/BL Lac boundary, or if a matched control sample of ordinary FSRQs shows the same crossing rate, the intermediate cluster does not actually track changing-look behaviour.","tokens_in":16249,"feed_emoji":"🔭","tokens_out":9174,"duration_ms":79864,"temperature":0.7,"pith_summary":"The paper claims that changing-look blazars form a coherent intermediate population between flat-spectrum radio quasars and BL Lac objects in the physical parameter space of gamma-ray, jet, and accretion-disk properties, and that this population can be recovered by unsupervised clustering. The authors apply Gaussian mixture modelling to 2250 blazars from the 4FGL-DR3 catalog, testing all 255 subsets of eight physical parameters. Four parameter combinations yield three-cluster solutions whose agreement with the 105 known changing-look blazars, measured by the adjusted Rand index, exceeds 0.61. The intersection of those four solutions contains 111 changing-look blazar candidates, including 44 previously reported sources and 67 new ones, 56 of which are labelled as FSRQs in the catalog. If the intermediate cluster is real, the changing-look phenomenon is not a handful of curiosities but a distinct, searchable class of active galactic nuclei.","feed_headline":"Clustering finds 67 new changing-look blazar candidates","feed_subtitle":"A three-way split of 4FGL blazars puts 111 sources in an intermediate group, 44 already known to switch spectral type.","key_machinery":"The central object is the three-component Gaussian mixture model fitted by the paper's mclust algorithm under the EVV covariance parameterization (ellipsoidal clusters of equal volume), with the number of components chosen by the Bayesian information criterion and the partition scored by the adjusted Rand index. The input features are eight physical parameters: gamma-ray photon index $\\Gamma_{\\rm ph}$, photon spectral slope $\\alpha_{\\rm ph}$, hardness ratios $HR_{34}$ and $HR_{45}$, Compton dominance $CD$, accretion disk luminosity $L_{\\rm disk}$, Eddington ratio $\\lambda = L_{\\rm disk}/L_{\\rm Edd}$, and redshift $z$. The machinery extracts the intermediate CLBC cluster and validates that it tracks the known changing-look blazars, rather than merely splitting the two standard classes.","core_discovery":"On the paper's own terms, the central discovery is that a Gaussian mixture model with three components, applied to the right subset of physical parameters, separates the 4FGL-DR3 blazar population into three groups that line up with BL Lacs, FSRQs, and changing-look blazar candidates (CLBCs), with the CLBC group sitting between the other two in every paired projection and in the dimension-reduced space. Four parameter subsets pass an adjusted Rand index threshold of 0.610 against the compiled catalog of known CLBs: ($\\alpha_{\\rm ph}$, CD, $L_{\\rm disk}/L_{\\rm Edd}$), (CD, $L_{\\rm disk}$, $L_{\\rm disk}/L_{\\rm Edd}$), ($\\Gamma_{\\rm ph}$, CD, $L_{\\rm disk}$, $L_{\\rm disk}/L_{\\rm Edd}$), and (HR45, CD, $L_{\\rm disk}$, $L_{\\rm disk}/L_{\\rm Edd}$). Taking the intersection of the CLBC predictions from these four subsets yields 111 candidates, of which 67 are new, and these candidates also occupy an intermediate position in WISE infrared color-color space.","pith_inferences":["Editorial inference: If the intermediate cluster is real, the 5 Å equivalent-width boundary used to separate FSRQs from BL Lacs is a binary projection of a continuous transition, and single-epoch spectroscopic classifications should be treated as probabilistic.","Editorial inference: The sharp drop in adjusted Rand index as more than four parameters are added suggests the CLB signal is concentrated in a narrow combination of features; future searches should test the four optimal subsets on independent samples before expanding the feature space.","Editorial inference: A straightforward extension would be to apply the fitted three-component model to the upcoming 4FGL-DR4 sample and ask whether the same intermediate group reappears with stable membership, which would test the reproducibility of the cluster structure.","Editorial inference: The clustering geometry implies CLBs sit near a critical Eddington-ratio threshold where broad-line region visibility switches; the model could be used to estimate that threshold from the boundary between the FSRQ and CLBC components."],"forward_implications":["The 67 new sources become the highest-priority targets for multi-epoch optical spectroscopy to confirm changing-look transitions.","Because 56 of the 67 new candidates are labelled FSRQs in 4FGL-DR3, existing FSRQ samples are likely to contain a hidden population of transition objects.","The four optimal parameter subsets identify the compact feature set (gamma-ray hardness plus Compton dominance and accretion-disk luminosity/Eddington ratio) that best separates the intermediate population.","The intermediate location of the candidates in WISE infrared color-color space, alongside their gamma-ray and disk parameters, indicates the intermediate character of CLBCs is not confined to a single waveband."],"supporting_citations":[{"why":"Establishes the eight physical parameters and the intermediate-position result that motivate this search.","marker":"Kang et al. 2024"},{"why":"Provides the TCLB catalog of 105 changing-look blazars used as the ground-truth labels for computing adjusted Rand index values.","marker":"Kang 2023"},{"why":"Supplies the Compton dominance, disk luminosity, and Eddington ratio measurements that appear in all four optimal parameter subsets.","marker":"Paliya et al. 2021"},{"why":"Provides the 4LAC-DR3 classifications and redshifts defining the 2250-source sample.","marker":"Ajello et al. 2022"},{"why":"Supplies 4FGL-DR3 photon indices, spectral slopes, and the energy fluxes used to compute hardness ratios.","marker":"Abdollahi et al. 2022"},{"why":"Documents the mclust Gaussian mixture modelling algorithm used for all clustering runs.","marker":"Scrucca et al. 2023"},{"why":"Defines the adjusted Rand index used to evaluate and select the four best clustering solutions.","marker":"Hubert & Arabie 1985"},{"why":"Provides the NbClust validation that independently assesses the number of clusters for the four optimal parameter subsets.","marker":"Charrad et al. 2014"},{"why":"Supplies the WISE blazar color-color sample used to show CLBCs also sit between FSRQs and BL Lacs in infrared space.","marker":"D'Abrusco et al. 2019"}],"fun_headline_variants":["67 new changing-look blazar candidates found via clustering","Blazar clustering yields 67 new changing-look candidates","Machine learning spots 67 changing-look blazar candidates","Clustering uncovers 67 new changing-look blazar candidates","67 new changing-look blazar candidates emerge from cluster analysis"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The result depends on the compiled catalog of 105 known changing-look blazars being correct and complete, since those same sources anchor both the choice of the eight parameters and the score used to pick the four best clustering solutions.","fun_headline_variants_meta":{"raw":{"variants":["67 new changing-look blazar candidates found via clustering","Blazar clustering yields 67 new changing-look candidates","Machine learning spots 67 changing-look blazar candidates","Clustering uncovers 67 new changing-look blazar candidates","67 new changing-look blazar candidates emerge from cluster analysis"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000368,"raw_usage":{"total_tokens":2063,"prompt_tokens":1120,"completion_tokens":943,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":736,"completion_tokens_details":{"reasoning_tokens":869}},"tokens_in":736,"tokens_out":943,"duration_ms":7850,"temperature":1.0,"reasoning_tokens":869,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T23:00:18.599420+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the 67 new candidate sources and obtain optical spectra at two or more epochs; if most show no change in broad emission-line equivalent width across the 5 Å FSRQ/BL Lac boundary, or if a matched control sample of ordinary FSRQs shows the same crossing rate, the intermediate cluster does not actually track changing-look behaviour.","supporting_citations":[{"cited_title":"2014, Journal of Statistical Software, 61, 1","cited_arxiv_id":null,"evidence_quote":"Provides the NbClust validation that independently assesses the number of clusters for the four optimal parameter subsets."}],"review_version":1}