{"id":"a9634e13-822c-42da-80c0-45d91edc5ab6","arxiv_id":"2506.12205","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Using eye-free morphological labels for roughly 14,000 CANDELS galaxies, the paper shows bulge-dominated galaxies split into a star-forming, disk-like population (G1) and an increasingly dominant quenched, massive population (G2) below z ~ 1.6.","lead":"This paper classifies about 14,000 distant galaxies by shape to chart how disks and bulges evolve over 10 billion years. It finds bulge-shaped galaxies divide into two groups, one still forming stars and one that has stopped, with the stopped group growing more common over time.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"G1 may be disk contamination: with disks outnumbering bulges ~1.7:1, the admitted pre-correction 18% disk-to-bulge misclassification would put ~30% interlopers in the bulge sample, enough to create the high-sSFR G1 component.","rationale":"The reader's weakest assumption is correct, and I find it is the load-bearing one. The distinctive observable of the paper is that bulge-dominated galaxies split into G1 (star-forming, low-mass, low-n, disk-like) and G2 (quenched, massive, high-n). Section 2.2.2 admits up to 18% of disks can be misclassified as bulge-dominated before correction; because disks outnumber bulges ~1.7:1 in the final sample, even a modest residual contamination places large numbers of true disks in the bulge catalog. The G1 component is defined by precisely those disk properties, and its weights in Table B4 are of the same order as the naive contamination fraction. Thus the GMM in Section 4.1 may be separating true quenched bulges from residual disk interlopers, not two real bulge populations. I also considered mass completeness and the imposed n=2 GMM as alternative concerns. Mass incompleteness (H<24 vs logM>=9) could affect the redshift trend of G2 prominence, but it does not by itself explain the existence of a low-redshift, high-sSFR, disk-like bulge component; the GMM non-uniqueness documented in Appendix C is a real limitation, but it would matter less if the input sample were clean. The paper's Section 5 caveats do not address residual classifier contamination, which is the key unvalidated link. The proposed confusion-matrix test is concrete and obtainable: it requires the authors to release the classifier and final classifications, then to subtract predicted interlopers and re-run the decomposition. If G1 persists after contamination subtraction, the two-track claim is supported; if not, the central claim reduces to a classifier artifact. Because this validation is achievable and the rest of the analysis is internally reasonable, CONDITIONAL remains the appropriate verdict until it is done.","tokens_in":25676,"tokens_out":10109,"duration_ms":127265,"concrete_test":"Apply the corrected Kolesnikov et al. (2025) classifier to a labeled sample with known morphologies spanning 0.2≤z≤2.4—e.g., FERENGI-redshifted local galaxies or the visual CANDELS classifications of Kartaltepe et al. (2015)—to produce a redshift-binned confusion matrix for the final classifier as actually used here. Then, in each redshift bin, subtract the expected disk→bulge interlopers from the bulge sample (or re-weight using the matrix) and re-run the two-component GMM of Section 4.1, recomputing G1/G2 weights and mean sSFR, mass, and Sérsic index. If G1's weight or mean properties shift by more than ~0.2 dex, or if the low-sSFR component no longer requires a high-sSFR companion, the two-track claim is not supported; if G1 survives contamination subtraction, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing assumption is that the \"bulge-dominated\" sample used in Sections 3 and 4 is genuinely bulge-dominated after the Kolesnikov et al. (2025) correction. Section 2.2.2 states that without correction up to 18% of disks could be misclassified as bulge-dominated; since the final sample contains 7,479 disks and 4,420 bulges, a residual misclassification near this level would put ~1,300 true disks into the bulge catalogue, i.e. ~30% of it. G1—the high-sSFR, lower-mass, lower-Sérsic component—has exactly the properties expected of such interlopers, and the G1 weights in Table B4 (0.31–0.52 below z<1.6) are of the same order as this contamination fraction. The final classifier is neither validated on an independent CANDELS-labeled sample nor accompanied by a confusion matrix in this paper. If G1 is mostly disk contamination, the headline \"two bulge evolutionary tracks\" becomes a mixture of true bulges and misclassified disks, and the merger-driven story built on G1/G2 loses its foundation.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper investigates the evolution of disk- and bulge-dominated galaxies using a CANDELS sample of ~14,000 galaxies selected with Hmag<=24, Mstellar>=1e9 Msun, and 0.2<=z<=2.4. Relying on the Kolesnikov et al. (2025) hybrid unsupervised-supervised morphological classification, the authors report that bulge-dominated galaxies develop a bimodal specific star formation rate (sSFR) distribution below z<1.6, whereas disks remain unimodal. A Gaussian mixture decomposition of the bulge sSFR distribution identifies two components: G1 (star-forming, lower mass, lower Sersic index) and G2 (quenched, more massive, higher Sersic index). The paper interprets these as two evolutionary tracks, with G2 formed through merger-driven transformations of massive disks, and presents mass-dependent evolution of morphological fractions as supporting evidence. The authors include caveats about SFR uncertainties and the indirect nature of the merger interpretation.","tokens_in":25936,"tokens_out":7536,"duration_ms":203171,"significance":"If the bimodality and the G1/G2 separation are real, this would be an observationally useful result: it would show that bulge-dominated galaxies are not a homogeneous population and that quenching and mass assembly may proceed along distinct paths. The paper has clear strengths: it uses an eye-free morphological classification, quantifies distribution shapes with detailed tables, includes an SFR-estimator comparison in Appendix A, and provides explicit caveats. However, the central claim is load-bearing on two under-supported choices: the adopted morphological classifier is not validated within this paper, and the two-component GMM is forced. Because G1 has exactly the properties expected of disk contaminants and the G1 weights are comparable to the stated misclassification rate, the uncertainty is not merely statistical. These issues need to be resolved before the main conclusion can be accepted.","major_comments":[{"comment":"The adopted morphological classification is not validated within this paper, and the admitted upper limit on disk-to-bulge misclassification is of the same order as the G1 component. The paper states that without correction up to 18% of disks could be misclassified as bulge-dominated, but it does not report the post-correction contamination rate or provide a confusion matrix for the final classifier on an independent CANDELS sample. With 7,479 disks and 4,420 bulges, an 18% contamination would place ~1,346 true disks into the bulge catalogue, i.e., ~30% of it. The G1 weights in Table B4 are 0.31-0.52 for z<1.6, comparable to this contamination fraction, and G1's properties (high sSFR, lower stellar mass, lower Sersic index) match the properties expected of disk contaminants. To support the claim that G1 is a genuine bulge subpopulation, the authors should quantify the post-correction purity (e.g., via cross-field validation, comparison with visual classifications, or a confusion matrix on a held-out set) and, ideally, repeat the GMM analysis on a bulge sample restricted to the highest-confidence classifications.","section":"2.2.2 / Table B4"},{"comment":"The choice of two Gaussian components is not used as a test of bimodality because the model selection is overridden. The text states that 'we decide to adopt n=2 for the GMM, irrespective of redshift' even when the AIC/BIC minimum lies at n!=2. The high-redshift bins (z>1.6) are explicitly noted as having weak bimodality, yet two components are still fitted and reported. To substantiate the claim that the sSFR distribution of bulge-dominated galaxies becomes bimodal below z<1.6, the paper should report the AIC/BIC values for n=1,2,3 in each redshift bin and apply a formal bimodality test (e.g., Hartigan's dip test or a calibrated likelihood-ratio). Without this, G1 and G2 are an assumed decomposition rather than an empirically supported separation.","section":"4.1 / Appendix C"},{"comment":"The sample is described as 'mass-complete', but the selection also includes Hmag<=24 and no redshift-dependent completeness analysis is presented. A fixed magnitude cut will remove low-mass galaxies at high redshift, so the Mstellar>=1e9 limit may not be complete over the full 0.2<=z<=2.4 range. This could bias the apparent redshift growth of the massive G2 component and the morphological fraction evolution in Section 3.3 (Table 1). The authors should show the 90% stellar-mass completeness limit as a function of redshift and restrict the evolution analysis to the mass-redshift region where the sample is complete.","section":"Abstract / Section 2"}],"minor_comments":[{"comment":"The phrase 'bulge-dominate' should be 'bulge-dominated'.","section":"2.2.2"},{"comment":"In the second bullet of the summary, 'galacies' should be 'galaxies'.","section":"5"},{"comment":"The labels 'BCSDSS', 'GVSDSS', and 'RSSDSS' are unclear; the caption should spell out that these are local-Universe boundaries from Trussler et al. (2020).","section":"Figure 7"},{"comment":"For the high-mass bulge-dominated bin, the fitted C_s = 1.26 +/- 0.5 exceeds the physical bound of unity for a fraction; consider a bounded fit or an explicit caveat.","section":"Table 1"},{"comment":"The sentence 'we consider only differences greater than 0.3 dex' is ambiguous; please clarify whether this refers to a threshold for trusting conclusions or a sample selection criterion.","section":"Appendix A"}],"recommendation":"major_revision","confidential_remarks":"The paper is well written and addresses a timely question, but the central claim depends on an external classifier that is not validated here and on a forced two-component GMM. The authors should be encouraged to add a validation appendix with a confusion matrix or independent cross-check, and to present the model-selection evidence for each redshift bin. The mass-completeness issue should also be addressed with a per-bin completeness analysis. These are fixable within the scope of the paper."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague, here is my read on arXiv:2506.12205. The genuinely new thing is applying the MEGG-based, eye-free classifier to CANDELS and claiming bulge-dominated galaxies split into two tracks, G1 and G2, with G2 growing and quenching below z~1.6. That is testable and interesting. The paper does a lot right: large sample, bootstrap uncertainties on the GMM parameters, a real Sersic-index/effective-radius comparison, and an appendix checking UV-corrected SFRs against UV+IR estimates. It also flags its own caveats, including that the merger story is indirect. Credit where due.\n\nThe soft spots are real, and the stress-test concern lands on reading the paper. Section 2.2.2 states that without correction, up to 18% of disks could be misclassified as bulge-dominated. Since disks outnumber bulges about 1.7:1, a residual error near that level would put roughly 30% interlopers into the bulge sample. G1 - high sSFR, lower mass, lower Sersic index - looks like exactly those interlopers, and its fitted weights (0.31-0.52 below z<1.6) are in the same ballpark as that contamination fraction. The paper does not show a confusion matrix for the corrected classifier or validate it on an independent CANDELS sample. The mass and Sersic offsets between G1 and G2 are consistent with contamination too, so they do not rescue the claim. This is the load-bearing problem.\n\nTwo lesser issues: the text says AIC/BIC favor n=2 below z<1.6, but Appendix C shows the choice is not unique at higher redshifts, and the authors impose n=2 everywhere. That is not fatal, but it should be reported more honestly. Also, the sample is selected by H_mag <= 24 plus a mass cut; calling it 'mass-complete' needs a completeness test that is not shown. The merger interpretation is indirect, but the authors acknowledge that, so I treat it as minor.\n\nOverall, the analysis is careful and transparent, but the headline claim depends on a classifier the reader cannot evaluate from this paper. I would send it to peer review, with a clear request: release the classifier and its validation, show a confusion matrix, demonstrate mass completeness, and test the bimodality with FIR-based SFRs where available. If G1 is mostly contamination, the two-track story loses its foundation; if it survives, it is an important result. Right now I would treat it as conditional.","headline":"A credible but conditional claim of two bulge evolutionary tracks; the missing classifier validation and plausible disk contamination keep me from endorsing it as is.","tokens_in":26516,"tokens_out":3187,"would_cite":false,"duration_ms":74776,"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":"Bulge-dominated galaxies split into two evolutionary tracks below $z \\sim 1.6$: G1, star-forming and disk-like at all redshifts, and G2, massive and quenched, built by merger-driven transformation of massive disks.","keywords":["galaxy evolution","galaxy morphology","star formation","high-redshift galaxies","bulge-dominated galaxies","CANDELS survey","Gaussian mixture decomposition","Sérsic index"],"falsifier":"Re-derive the disk/bulge labels for the same galaxies with an independent method — spatially resolved stellar kinematics from integral-field spectroscopy at $0.5 < z < 1.5$, or a classifier built on different morphological indices — and re-run the Gaussian mixture decomposition on the bulge-dominated subset. If the two-component sSFR split and the massive, high-Sérsic G2 population disappear under the alternative labels, the classification was creating the signal; if they persist, the two-track picture holds.","tokens_in":25434,"feed_emoji":"🌌","tokens_out":15655,"duration_ms":150457,"temperature":0.7,"pith_summary":"Using a mass-complete sample of about 14,000 galaxies from the CANDELS survey spanning $0.2 \\le z \\le 2.4$, the paper claims that disk and bulge-dominated galaxies begin with nearly identical specific star formation rates at $z \\sim 2.4$ but diverge sharply below $z < 1.6$. The divergence takes a specific form: the sSFR distribution of bulge-dominated galaxies becomes bimodal, splitting into G1, a long-lived star-forming population with disk-like properties, and G2, a quenched population that is about 0.6 dex more massive, about 0.9 dex less active in star formation, and increasingly dominant toward the present. The paper argues the split is driven by stellar mass assembly through major mergers of massive disks, not by smooth quenching alone, and that the transformation is strongly mass-dependent. If correct, bulge-dominated galaxies are not one population with one formation channel but two families following divergent physical pathways.","feed_headline":"Bulge galaxies split into two evolutionary tracks by z<1.6","feed_subtitle":"One track keeps forming stars; the other grows massive and quenches, reshaping how bulges form.","key_machinery":"The argument rides on two tools. The first is an eye-free, redshift-binned morphological classification built from the MEGG non-parametric indices — second moment of light $M_{20}$, Shannon entropy, Gini coefficient, and gradient field asymmetry — grouped by a self-organizing map and labeled by an ensemble of convolutional neural networks, so that galaxies count as disk- or bulge-dominated without visual inspection or parametric light-profile assumptions. The second is a Gaussian mixture decomposition (Bayesian and frequentist, with the component number chosen by AIC/BIC) applied to the sSFR distribution of bulge-dominated galaxies in each redshift bin, which isolates the two tracks G1 and G2. Sérsic structural parameters (effective radius $R_e$, Sérsic index $n_s$) from multi-band profile fits and SED-based stellar masses and star formation rates from the CANDELS catalogs supply the physical properties that give the two components distinct identities.","core_discovery":"The central claim is that bulge-dominated galaxies are heterogeneous: at $z < 1.6$ their sSFR distribution separates into two Gaussian components with a consistent mean offset of $\\langle\\Delta\\mu\\rangle = 1.31 \\pm 0.19$ dex. G1 systems remain on the blue cloud of the star-forming main sequence at every redshift, with Sérsic indices and effective radii close to disk values, indicating sustained star formation despite a bulge-dominated morphology. G2 systems are more massive ($\\Delta\\log M_*/M_\\odot = 0.61 \\pm 0.16$ dex), have lower star formation rates ($\\Delta\\log\\mathrm{SFR} = 0.92 \\pm 0.26$ dex), higher Sérsic indices ($\\Delta n_s = 0.99 \\pm 0.35$) at similar effective radii, and migrate from the blue cloud into the green valley and red sequence as redshift declines; their mean stellar mass changes little, arguing against growth through in-situ star formation. The paper reads this as merger-driven transformation of massive disks into quenched, centrally concentrated remnants: between $z = 2.4$ and $z = 0.2$ the fraction of massive ($\\log M_*/M_\\odot > 10.5$) disks falls from 55% to 5% while the bulge-dominated fraction rises from 25% to 90%, with the steepest trends in exactly the mass bin where G2 grows.","pith_inferences":["If G1 truly persists as a star-forming, bulge-dominated population over the full 10 Gyr baseline, it may trace a formation channel — bulge growth through disk instabilities or pseudo-bulge assembly without quenching — that the merger narrative for ellipticals does not cover.","The paper's implication that most quenching happens recently ($z < 0.3$) is testable with local data: applying the same UV-based SFR calibration to low-redshift spectroscopic samples should show the disk-to-bulge SFR gap widening from about 0.3 dex toward the roughly 1 dex local value over the past ~2 Gyr.","The merger interpretation carries an independent check: if G2 is merger-built, independent merger indicators such as close-pair fractions or morphological disturbance rates at $z \\sim 1$–$2$ should rise in lockstep with the G2 fraction."],"forward_implications":["Below $z < 1.6$, bulge-dominated galaxies are bimodal in both sSFR and stellar mass, so analyses that treat them as a single population average over two distinct evolutionary states.","The most massive disks ($\\log M_*/M_\\odot > 10.5$) drop from 55% to 5% of the morphological mix between $z = 2.4$ and $z = 0.2$ while massive bulge-dominated systems rise from 25% to 90%, implying a mass-dependent transformation that reshapes the high-mass end of the galaxy population.","The SFR gap between disks and bulge-dominated galaxies is only about 0.3 dex at $z \\sim 0.2$, so most quenching of bulge-dominated systems must happen below $z < 0.3$, within roughly 2 Gyr.","G2's nearly constant mean stellar mass despite growing prominence, together with its higher Sérsic index at comparable effective radius, points to assembly by dissipative major mergers rather than in-situ star formation.","Intermediate-mass bins ($9.5 \\le \\log M_*/M_\\odot < 10$) show nearly static morphological fractions, which can hide the transformation and explain why some surveys report weak morphological evolution."],"supporting_citations":[{"why":"Supplies the hybrid SOM-CNN, redshift-binned morphological classifier that assigns every galaxy its disk or bulge-dominated label.","marker":"Kolesnikov et al. (2025)"},{"why":"Establishes the MEGG non-parametric indices as the feature set that separates elliptical and spiral galaxies, on which the classifier is built.","marker":"Kolesnikov et al. (2024)"},{"why":"Provides the CANDELS multi-wavelength catalog with SED-based stellar masses and the UV-corrected star formation rates used throughout the paper.","marker":"Barro et al. (2019)"},{"why":"Provides the stellar mass estimates that set the mass-complete sample limit at $10^9\\,M_\\odot$.","marker":"Santini et al. (2015)"},{"why":"Supplies the multi-band Sérsic fits (effective radius and Sérsic index) used to compare the structure of G1 and G2.","marker":"Nedkova et al. (2024)"},{"why":"Provides the merger-driven bulge formation and quenching framework used to interpret G2 as a merger product.","marker":"Hopkins et al. (2009)"},{"why":"Supports the claim that mergers build stellar mass in bulge-dominated systems at the expense of massive disks.","marker":"Rodriguez-Gomez et al. (2016)"},{"why":"Offers the fast- and slow-track quenching picture that frames two distinct pathways to quiescence.","marker":"Pandya et al. (2017)"}],"fun_headline_variants":["Bulge galaxies split into two tracks by redshift 1.6","Bulges diverge: one star-forming, one quenched and massive","Bulge bimodality: quenched path tied to mass and mergers","Massive disks transform into quenched bulges over time","Bulge galaxies not homogeneous: two evolutionary paths found"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The redshift-binned morphological classifier adopted from the companion catalog must be accurate enough that the disk/bulge labels do not themselves manufacture the observed bimodality; the paper reports that up to 18% of disks could be misclassified as bulge-dominated without correction, and the corrected classifier is taken as input rather than independently validated here.","fun_headline_variants_meta":{"raw":{"variants":["Bulge galaxies split into two tracks by redshift 1.6","Bulges diverge: one star-forming, one quenched and massive","Bulge bimodality: quenched path tied to mass and mergers","Massive disks transform into quenched bulges over time","Bulge galaxies not homogeneous: two evolutionary paths found"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000279,"raw_usage":{"total_tokens":1761,"prompt_tokens":1156,"completion_tokens":605,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":772,"completion_tokens_details":{"reasoning_tokens":517}},"tokens_in":772,"tokens_out":605,"duration_ms":7256,"temperature":1.0,"reasoning_tokens":517,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T00:56:45.910524+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-derive the disk/bulge labels for the same galaxies with an independent method — spatially resolved stellar kinematics from integral-field spectroscopy at $0.5 < z < 1.5$, or a classifier built on different morphological indices — and re-run the Gaussian mixture decomposition on the bulge-dominated subset. If the two-component sSFR split and the massive, high-Sérsic G2 population disappear under the alternative labels, the classification was creating the signal; if they persist, the two-track picture holds.","supporting_citations":[{"cited_title":"M., de Carvalho R","cited_arxiv_id":null,"evidence_quote":"Supplies the hybrid SOM-CNN, redshift-binned morphological classifier that assigns every galaxy its disk or bulge-dominated label."}],"review_version":1}