{"id":"d1867905-cfea-4582-b904-ee740b633d37","arxiv_id":"2608.06604","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A hybrid self-supervised and consensus-supervised model yields calibrated rotation periods for 148,746 Kepler main-sequence stars and identifies bimodal signals where the longer mode is the true rotation.","lead":"The Maunder, a machine learning pipeline trained partly on labels agreed by several existing catalogs, produces rotation periods for 148,746 Kepler main-sequence stars, with calibrated uncertainties for each star. It finds that 21.5% of these stars show two competing rotation signals and uses independent spectroscopic data to show the longer signal is usually the true spin, which would mean many classical rotation periods are alias artifacts.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The long-mode conclusion assumes isotropic inclinations for the 390 bimodals, but the paper itself argues low-coherence stars are preferentially low-inclination; if bimodals share that bias, the short mode may fit with an anisotropic prior.","rationale":"The reader's weakest assumption identifies exactly this: the v sin i forward model is calibrated on unambiguous unimodal stars and applied to only 390 bimodals under an isotropic-inclination assumption. My read agrees and sharpens it with an internal tension: the paper itself argues in Section 4.2 that low-coherence stars are preferentially low-inclination, and Section 4.3 shows bimodals are disproportionately low-coherence. If so, the isotropic prior is invalid for the very subsample used to choose the long mode, and the short mode could be rescued by an anisotropic prior. That would not invalidate the catalog as a consensus-period product, but it would remove the headline claim that ~26,000 stars are systematically mis-assigned. The paper has genuine independent support: held-out consensus RMSE of 2.36 days, calibrated conformal intervals, recovery of gyrochronology sequences, the Kraft-break velocity steps, and the metallicity–rotation trend all argue that the catalog is useful. But these validations are largely on or near the consensus subset and do not test the bimodal adjudication. The proposed test — freeing the inclination prior — is the single check that would settle whether the long-mode conclusion is an artifact of the isotropy assumption. Until that is done, conditional acceptance is the right posture; no verdict change is needed.","tokens_in":25177,"tokens_out":4244,"duration_ms":43610,"concrete_test":"Re-run the Section 4.3 hierarchical forward model on the N=390 distinct bimodals with the inclination prior as a free parameter — e.g., a one- or two-parameter beta density on sin i, or a free fraction in a low-inclination component — while keeping the macroturbulence and measurement-floor parameters fixed at their control-calibrated values. Compare long-mode versus short-mode fits by likelihood ratio or Bayesian evidence marginalized over the inclination prior. If the short mode fits with an inclination prior that is consistent with the low-amplitude bias documented in Section 4.2 for bimodals, the headline claim fails. As a sanity check, apply the same free-prior fit to the N≃2,790 unimodal control to verify that the parameters are identifiable and that the control population is consistent with isotropy.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim — that the longer mode is the true rotation for distinct bimodals — rests entirely on the hierarchical v sin i forward model of Section 4.3. That model assumes 'over an isotropically-oriented population, cos i ∼ U(0,1)', calibrates its macroturbulence/measurement-floor parameters on N≃2,790 unimodal control stars, and then applies the frozen model to N=390 distinct bimodals. The short mode is rejected because it predicts too many nearly pole-on stars under the isotropic prior. But Section 4.2 independently argues, from photometric amplitude and planet-host comparisons, that low-coherence populations are preferentially low-inclination, noting the strong correlation (Spearman ρ=0.79) between low ACF coherence and low amplitude. Bimodals are disproportionately low-coherence: the bimodal fraction rises monotonically as the Basri et al. spot-lifetime estimator decreases (Figure 10). If the bimodal population inherits this low-inclination bias, then the isotropic prior is violated for exactly the sample used to adjudicate the two modes. The short mode's 'unphysical excess of rapid rotators' is precisely a deficit of inferred sini, which is what a low-inclination bias would produce. Thus the long mode's good fit (KSD=0.10) may be a coincidence of applying an isotropic model to a biased sample, and the conclusion that ~26,000 stars are mis-assigned by single-pass periodograms is not secured. The control calibration does not test this, because the control stars are unimodal and more coherent, i.e., presumably more isotropic. This is an internal tension, not merely a disagreement with consensus.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents The Maunder, a machine-learning pipeline that predicts rotation periods for 148,746 main-sequence stars in the Kepler field. The model uses a hybrid objective: a self-supervised joint-embedding loss applied to all light curves, plus a supervised quantile-regression loss trained on cross-catalog consensus labels from four period catalogs. Rolling-window inference yields per-star period quantiles and identifies 31,953 bimodal stars (21.5%). Using APOGEE vsini and Berger et al. radii, the authors build a hierarchical forward model and argue that for distinct (non-harmonic) bimodals the longer period is the true rotation, implying that classical single-pass periodograms systematically lock onto shorter aliases for roughly 26,000 stars. A CI-based filter produces a 'highly reliable subset' of 119,428 stars, which is then used to recover gyrochronology sequences, a metallicity-rotation trend at fixed mass, equatorial velocities across the Kraft break, and candidate hierarchical triples among binaries. The paper also releases a per-star and a window-level catalog with quantiles, bimodality flags, and uncertainty diagnostics.","tokens_in":25526,"tokens_out":7968,"duration_ms":68910,"significance":"If the central long-mode claim holds, the paper identifies a population-level systematic error in previous rotation catalogs and provides the largest Kepler main-sequence rotation catalog with per-star calibrated uncertainties. The catalog design is thoughtful: the release of both per-star and window-level predictions, the explicit bimodality diagnostics, and the candid discussion of limitations (no direct error metric for the ambiguous majority, conformal coverage only on exchangeable stars, harmonic-aliased failures that survive the CI cut) are all strengths. However, the long-mode conclusion rests on an isotropy assumption that conflicts with the paper's own evidence that low-coherence (and hence bimodal) stars are preferentially low-inclination; this is load-bearing. The circular validation of the filtered catalog against the same catalogs that produced the training labels further limits the strength of the reliability claims. The catalog could still be a valuable resource even if the mode-adjudication claim needs revision, but the current evidence does not secure the headline result.","major_comments":[{"comment":"The paper's own analysis implies that the bimodal sample violates the isotropy assumption of the vsini forward model. In Sec. 4.2, the authors argue from photometric amplitudes and planet-host fractions that low-coherence stars are preferentially low-inclination, and in Fig. 10 bimodals are shown to be disproportionately low-coherence. The forward model in Fig. 9 instead assumes cos i ~ U(0,1) for the bimodal population. If bimodals inherit a low-inclination bias, the short-mode hypothesis predicts an excess of low sini values, which the isotropic model interprets as an 'unphysical excess of rapid rotators' (D=0.48); the long mode's good fit (KSD=0.10) may then be an artifact of applying an isotropic model to a biased sample. The control calibration on N~2,790 unimodal stars does not test this, because those stars are not representative of the bimodal population. The headline claim that ~26,000 stars are mis-assigned by single-pass periodograms is therefore not secured by the presented evidence.","section":"Sec. 4.2 and Sec. 4.3, Fig. 9"},{"comment":"The hierarchical forward model used to adjudicate the two modes is not described in enough detail to be reproduced. The text says the model follows Masuda & Winn (2020) and the appendix lists calibrated values (sigma_0 = 1.02, v_mac = 2.53, f = 0.05), but the likelihood, the prior on inclination, the treatment of the 1.5 km/s detection truncation, and the procedure for calibrating the noise parameters on the unimodal control are not given. Because this model is the sole quantitative evidence for the paper's central claim, the missing equations and algorithmic details are a load-bearing gap.","section":"Sec. 4.3 and Appendix 7.2"},{"comment":"The reliability of the filtered subset is validated by agreement with the same catalogs that supplied the consensus training labels. The period distribution after the CI80/Prot<0.4 cut is compared with McQuillan et al. (2014) and Santos et al. (2021), both of which are among the four input catalogs used to build the labels. The held-out test set is likewise drawn from the same consensus construction, so the RMSE and coverage numbers do not provide independent evidence for the unlabeled majority. The paper correctly states in the limitations that no direct error metric exists for the ambiguous majority, but the specific claim in Sec. 4.4 that the filtered catalog 'agrees with previous catalogs' should be re-framed as a consistency check rather than independent validation.","section":"Sec. 4.4, Fig. 12"},{"comment":"The extrapolation from N=390 distinct bimodals with APOGEE vsini to all 25,357 distinct bimodals is not justified. APOGEE targets are subject to selection effects (brightness, temperature, and survey footprint), and no test is presented that the 390-star subset is representative of the full distinct-bimodal population in period, amplitude, coherence, or inclination. A selection-bias analysis or a demonstration that the long-mode preference is homogeneous across the parameter space of bimodals is needed before the population-level claim is supported.","section":"Sec. 4.3, Fig. 9"}],"minor_comments":[{"comment":"The phrase 'pointing on the role of metallicity' should be 'pointing to the role of metallicity'; the same wording appears in the main text.","section":"Abstract"},{"comment":"The definition of the consensus label is clear, but it is not stated whether the final average period is computed in linear or logarithmic space; please state this explicitly.","section":"Sec. 2.1"},{"comment":"The phrase 'they didn't use the same training labels' is informal for a journal paper; use 'they did not use'.","section":"Sec. 4.1"},{"comment":"The term 'planet host stars' is undefined in the caption; please specify the source of the planet-host sample (e.g., confirmed/candidate Kepler planets) and the cross-match used.","section":"Fig. 6 caption"},{"comment":"The harmonic-alias classification uses the threshold |Delta log10 P - log10 2| < 0.06 dex; please add a sentence justifying this tolerance.","section":"Sec. 4.3"},{"comment":"The catalog is stated to be 'available online upon publication'; for a catalog paper, please provide the expected archive/DOI or an anonymous access link for reviewers.","section":"Sec. 5"},{"comment":"The age normalization is taken from A. Sussholz et al. (2026), an arXiv preprint; please ensure the description of TAMS(M) and the YREC grid interpolation is self-contained or include a reference to the published version.","section":"Sec. 4.6"},{"comment":"There are minor typographical and formatting issues, including the missing space in 'The Maunderprovides' in the abstract and inconsistent use of 'vsini' versus 'v sin i' in equations and text.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The paper is well-suited to an astronomy journal and the release of a large Kepler rotation catalog with per-star uncertainty metrics is potentially valuable. However, the central claim about the long mode being the true rotation is not supported by the current evidence: the vsini forward model assumes isotropic inclinations for a population the paper itself argues is low-inclination. The authors should be asked to re-run the mode-adjudication with an inclination model that accounts for the low-coherence bias, or to obtain independent validation (e.g., from stars with known inclinations or with alternative spectroscopic constraints). The vsini model details must also be provided. The catalog itself may remain useful even if the long-mode claim is weakened, but the current manuscript overstates the certainty of the 'systematic failure mode' result. The circular validation of the filtered subset should be transparently reframed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. First, The Maunder is a serious piece of work: 148,746 main-sequence Kepler periods, a hybrid consensus-plus-self-supervised training scheme, conformal quantile intervals, and a transparent limitations section. Second, the central new physical claim — that for distinct bimodals the long period is the true rotation, implicating ~26,000 stars — is not secured, and the reason is an internal tension the authors don't address.\n\nThe vsini adjudication assumes an isotropically oriented population (cos i uniform). The short mode fails that test because it implies an excess of near-pole-on stars. But in Section 4.2 the authors argue, using amplitude and planet-host comparisons, that low-coherence stars are preferentially low-inclination, and note the strong amplitude–coherence correlation (ρ = 0.79). Bimodals are disproportionately low-coherence (Fig. 10). So the very population being adjudicated is the one the authors expect to violate isotropy. The short mode's 'excess of rapid rotators' is exactly the signature a pole-on bias would produce. The control calibration on ~2,790 unimodal stars doesn't fix this, because unimodal stars are more coherent and plausibly closer to isotropic. So the long mode's good KSD fit (0.10) may be a coincidence of applying an isotropic model to a biased sample. The stress-test note is right; this is a real, load-bearing weakness.\n\nWhere the paper deserves credit: the catalog itself is a potentially major resource, even if the long-mode rule is wrong for some of the 25k distinct bimodals. The per-window catalog and the two uncertainty metrics are good ideas. The paper honestly flags that accuracy is only measured on the consensus subset, that harmonic aliases survive the recommended cut, and that the Basri spot-lifetime extrapolation is uncalibrated. The recovery of known gyrochronology, Kraft-break, metallicity, and binary trends is reassuring.\n\nThe circularity in validation — the reliability filter is tested against the same catalogs that supplied the labels — is real but acknowledged; it limits the 'highly reliable subset' claim without invalidating the catalog. The small vsini sample (390 of 25,357 distinct bimodals) also makes the extrapolation to the full population fragile.\n\nWho should read this: stellar rotation and gyrochronology people, especially those who want a large catalog to test selection effects. It deserves a serious referee, but a major revision should redo the vsini analysis with an anisotropic inclination model — or at least test whether the conclusion survives when bimodals are allowed to be low-inclination. As it stands, the headline claim is a hypothesis, not a demonstration.","headline":"A large, carefully built Kepler rotation catalog, but the headline claim that ~26,000 bimodal stars were mis-assigned rests on an inclination prior that the paper itself undermines.","tokens_in":26108,"tokens_out":5561,"would_cite":true,"duration_ms":52943,"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 Maunder catalog shows that for Kepler main-sequence stars whose light curves carry two distinct rotation-like signals, spectroscopic v sin i measurements identify the longer period as the true rotation, implying that classical…","keywords":["stellar rotation periods","Kepler main-sequence stars","machine learning period estimation","joint-embedding self-supervised learning","bimodal rotation signals","APOGEE v sin i","magnetic braking","gyrochronology"],"falsifier":"The most direct settlement is to measure rotation periods for a sample of distinct bimodals by an independent technique, such as high-resolution time-series spectroscopy of spot-induced line-profile variations: if a substantial fraction rotate at the short mode, the central claim fails. A cheaper calculation is to refit the $v \\sin i$ noise model to the bimodal sample itself and check whether the long-mode advantage ($D = 0.10$ versus $0.48$) survives down to the control calibration residual of $0.12$.","tokens_in":2341,"feed_emoji":"🔄","tokens_out":5612,"duration_ms":125213,"temperature":0.7,"pith_summary":"The paper builds a machine-learning pipeline, The Maunder, to measure rotation periods for 148,746 main-sequence stars in the Kepler field, and releases a catalog of 119,428 periods it considers reliable. Its central discovery comes from sliding-window inference: for 21.5% of the sample the model does not settle on one period but alternates between two, and a comparison with APOGEE projected rotation velocities shows that for non-harmonic double signals the longer period is the true spin, while the shorter one is an alias. The authors conclude that roughly 26,000 stars in earlier single-pass periodogram analyses have been systematically assigned the shorter, wrong rotation period. If correct, this corrects about 26,000 mis-perioded stars and anchors population studies of gyrochronology, magnetic braking, and binary synchronization on the largest Kepler main-sequence sample with per-star calibrated uncertainties.","feed_headline":"Longer of two signals is the true rotation for 26,000 stars","feed_subtitle":"Spectroscopically anchored Kepler catalog shows classical period searches pick a shorter alias for 25,000 stars.","key_machinery":"The load-bearing mechanism is rolling-window inference plus adjudication by projected rotation velocity. The model predicts period quantiles on 450-day windows slid across each roughly four-year light curve with a 90-day stride; a 2-means split in $\\log_{10} P$ classifies a star as bimodal when two period clusters are separated by at least 0.20 dex with a clean partition, and splits near a 2:1 ratio are set aside as harmonic aliases. For distinct bimodals, a hierarchical forward model of APOGEE $v \\sin i$ — built following the Masuda–Winn approach with isotropic inclinations, a 1.5 km/s detection truncation, and a macroturbulence-plus-measurement floor calibrated on a ~2,790-star unimodal control — decides which mode is the rotation: the long mode reproduces the observed $v \\sin i$ distribution, while the short mode predicts an unphysical excess of rapid rotators. The six-channel input (two flux normalizations, two activity proxies windowed by a scaffold period, plus autocorrelation and Lomb–Scargle channels) and the hybrid objective (a joint-embedding self-supervised loss on all stars combined with a conformalized quantile-regression loss on 41,650 cross-catalog consensus labels) supply the calibrated per-star intervals that make the catalog usable.","core_discovery":"The central claim is that rolling-window period inference exposes a bimodality that single-pass period searches cannot see, and that the two modes are not both physical: for distinctly separated, non-harmonic bimodals (25,357 of the 31,953 bimodal stars), the longer mode is the true rotation period, and the shorter mode is an alias. The evidence is a hierarchical forward model of APOGEE $v \\sin i$ distributions: assuming isotropic inclinations and a noise floor calibrated on about 2,790 unambiguous-period control stars, adopting the long mode as the rotation reproduces the observed $v \\sin i$ of the 390 distinct bimodals with APOGEE data as well as the control does (KS distance 0.10 versus 0.12), while the short mode predicts a large overabundance of rapid rotators ($D = 0.48$). The paper adopts the long mode for distinct bimodals in the released catalog, flags 2:1 harmonic splits as ambiguous because $v \\sin i$ cannot resolve them, and publishes both candidate modes for every star so users can revert the choice. With a confidence-interval cut the catalog yields 119,428 stars, and the paper uses it to recover the metallicity dependence of rotation at fixed mass, the jump in equatorial velocity and specific angular momentum across the Kraft break, the empirical gyrochronology sequences, and a candidate population of hierarchical triples among synchronized binaries.","pith_inferences":["The mis-assignment rate should be concentrated rather than uniform: because the paper's own spot-lifetime analysis ties bimodality to short-lived spots, earlier catalogs' fast-rotator populations made of low-coherence, low-amplitude stars are the most likely to harbor alias-contaminated periods, and a star-by-star comparison of classical periods against the long modes would show where the damage i","The $v \\sin i$ adjudication is statistical, resting on only 390 stars with APOGEE data; as larger spectroscopic surveys provide $v \\sin i$ for the remaining ~25,000 distinct bimodals, the long-mode rule could gain per-star verification or reveal sub-populations, such as genuinely mode-switching stars, that violate it.","The hierarchical-triple interpretation of the wide-orbit, short-rotation regime predicts that the short-period component is itself a close binary and that the wide companion should be visible in radial velocities; a handful of RV epochs for regime-B stars would settle the interpretation.","Even if the $v \\sin i$ adjudication were weakened, the catalog would not lose all value: the consensus-grounded periods and calibrated intervals would still support relative population comparisons, and only the period-source choice for distinct bimodals would need revisiting."],"forward_implications":["About 26,000 main-sequence stars with distinct bimodal signals carry rotation periods that classical single-pass periodogram analyses systematically set to a shorter alias; the catalog instead adopts the long mode.","The confidence-filtered catalog of 119,428 stars, with calibrated per-star uncertainties, provides rotation periods for the largest Kepler main-sequence population, supporting population-level gyrochronology, spin-orbit, and magnetic-braking studies.","At fixed stellar mass above 0.85 $M_{\\odot}$, median rotation period increases monotonically with metallicity (for example from 14.0 to 21.5 days at 1.0–1.15 $M_{\\odot}$), a trend opposite to the age–metallicity relation, indicating metallicity-dependent magnetic braking.","Equatorial velocities and specific angular momenta traced directly from periods and radii rise steeply across the Kraft break, from roughly 10–20 km/s at 1.3 $M_{\\odot}$ to about 100 km/s at 1.6 $M_{\\odot}$.","A regime of short rotation periods at wide orbital separations among known binaries is interpreted as hierarchical triples, with elevated astrometric noise and one confirmed triple system (KID 6525196) supporting the identification."],"supporting_citations":[{"why":"Supplies one of the four reference catalogs whose agreement defines the supervised consensus rotation labels.","marker":"A. McQuillan et al. (2014)"},{"why":"Supplies a second consensus catalog for the supervised labels and a baseline for the filtered period distribution.","marker":"A. R. G. Santos et al. (2021)"},{"why":"Supplies a third consensus catalog and, for the showcase star KID 892376, the long-period mode that matches the model's long cluster.","marker":"T. Reinhold et al. (2023)"},{"why":"Supplies the fourth consensus catalog and the prior simulation-trained pipeline whose disagreements highlight the simulation-to-reality gap.","marker":"I. Kamai & H. B. Perets (2025a)"},{"why":"Supplies the hierarchical forward-model structure used to adjudicate long versus short mode with APOGEE v sin i.","marker":"K. Masuda & J. N. Winn (2020)"},{"why":"Supplies the APOGEE DR17 v sin i measurements that carry the mode adjudication.","marker":"Abdurro'uf et al. (2022)"},{"why":"Supplies the radii, masses, and metallicities used to convert period to implied inclination and to run the mass and metallicity analyses.","marker":"T. A. Berger et al. (2020)"},{"why":"Supplies the conformalized quantile regression that gives the per-star predictive intervals their calibrated coverage.","marker":"Y. Romano et al. (2019)"},{"why":"Supplies the coherence parameter and spot-lifetime estimator used to explore why low-coherence stars become bimodal.","marker":"G. Basri et al. (2022)"}],"fun_headline_variants":["AI finds rotation for 148k Kepler stars, exposes alias trap","Long rotation wins for 25k bimodal stars in Kepler catalog","Bimodal rotation? Longer period is the real one, AI shows","Kepler rotation catalog: 119k reliable periods from 148k stars","AI exposes alias failure in classical rotation searches for Kepler stars"],"cache_read_input_tokens":28032,"weakest_assumption_plain":"The conclusion that the long mode is the true rotation for roughly 26,000 stars rests on a single noise model for APOGEE projected rotation velocities, calibrated on 2,790 unambiguous stars and applied unchanged to 390 bimodal stars; if those two populations differ in inclination distribution, detection floor, or line-broadening behavior, the long-mode verdict is not established.","fun_headline_variants_meta":{"raw":{"variants":["AI finds rotation for 148k Kepler stars, exposes alias trap","Long rotation wins for 25k bimodal stars in Kepler catalog","Bimodal rotation? Longer period is the real one, AI shows","Kepler rotation catalog: 119k reliable periods from 148k stars","AI exposes alias failure in classical rotation searches for Kepler stars"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000607,"raw_usage":{"total_tokens":2927,"prompt_tokens":1143,"completion_tokens":1784,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":759,"completion_tokens_details":{"reasoning_tokens":1692}},"tokens_in":759,"tokens_out":1784,"duration_ms":10925,"temperature":1.0,"reasoning_tokens":1692,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T04:16:51.890036+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"The most direct settlement is to measure rotation periods for a sample of distinct bimodals by an independent technique, such as high-resolution time-series spectroscopy of spot-induced line-profile variations: if a substantial fraction rotate at the short mode, the central claim fails. A cheaper calculation is to refit the $v \\sin i$ noise model to the bimodal sample itself and check whether the long-mode advantage ($D = 0.10$ versus $0.48$) survives down to the control calibration residual of $0.12$.","supporting_citations":[],"review_version":1}