{"id":"c28d080e-337c-49aa-b2d4-f812feeea921","arxiv_id":"2504.20327","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Five known nearby halo substructures are re-detected with HDBSCAN* clustering in DESI Year 1 data, and DESI metallicities confirm three of them as chemically distinct.","lead":"Using DESI Year 1 spectra of 138,661 nearby halo stars, the authors identify five kinematic groups and match each one to a previously known Milky Way substructure, including Helmi streams and Sequoia. The study mainly confirms known structures in a new dataset and adds DESI metallicity measurements, rather than reporting new discoveries.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The validation claim rests on low-stability clusters B–E; the stability metric in §3.2 is uncalibrated against a null model, so it is unknown whether 23–40% IoM recovery means fragile real structures or chance fluctuations.","rationale":"The paper is honest and technically competent, and I do not find a reason to doubt the authors' good faith. The strongest claim—that the five HDBSCAN* groups confirm known substructures and validate the method—depends on the clusters being genuine. The most insecure link is the small, low-stability clusters B–E: four of five have IoM recovery under measurement perturbation below 41%, and Cluster C is not independently detected in velocity space and is not chemically distinct. The stability metric has no calibration against null data, and the mock tuning reports completeness and purity for injected streams but not a false-positive rate for the adopted hyperparameters. This is exactly the missing test needed before using HDBSCAN* for blind discovery. The paper itself flags related limitations in §4.3, which supports a conditional rather than an accept verdict. A null-realization and injection-recovery test would settle whether the low stability rates are a property of genuine small clusters under DESI errors or evidence that B–E are marginal. I therefore keep the reader's CONDITIONAL verdict pending that check.","tokens_in":43447,"tokens_out":5508,"duration_ms":62894,"concrete_test":"Run the same HDBSCAN* pipeline with the same hyperparameters on an ensemble of null realizations of the 138,861-star halo subset in which the physical association between each star's DESI radial velocity and its Gaia astrometry is destroyed by shuffling (or by sampling actions/velocities from a multivariate Gaussian fit to the smooth halo), preserving the marginal distributions and selection function. Count clusters with ≥20 members in the same IoM regions. Independently calibrate the §3.2 stability statistic by injecting mock streams of 20–50 stars with realistic DESI distance and velocity uncertainties into the real halo subset and repeating the 200-iteration perturbation.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 5's claim that the detections 'confirm the presence of known substructures' and validate HDBSCAN* requires the clusters to be genuine overdensities. The weakest evidence is for Clusters B–E: the perturbation test in §3.2 (Table 4) yields only 23.5% (B), 27.0% (C), 25.0% (D), and 40.5% (E) stability in IoM space. The stability definition—a candidate must contain >50% of its members in the original 99% ellipsoid after re-clustering—is threshold-based and uncalibrated: the mocks in §2.4.1 use Gaussian thick-disc/GSE/other components with an assumed 50% GSE fraction and report completeness/purity for injected streams, but never a false-positive rate for the adopted hyperparameters on a smooth-halo null. Cluster C is detected only in IoM space and is not chemically distinct from background (KS p=0.065). Without a null-model false-positive rate, we cannot tell whether 23–40% recovery is normal for genuine 20–50 star clusters under realistic distance errors or whether these are marginal density fluctuations. Since the paper proposes HDBSCAN* for future blind discovery, this missing calibration is the load-bearing gap.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper applies the unsupervised clustering algorithm HDBSCAN* to a subset of 138,861 nearby (d < 5 kpc) halo stars from the DESI Milky Way Survey Year 1 catalogue, searching for kinematic substructures in both Integrals of Motion space (E_tot, L_z, log J_r, log J_z) and Galactocentric cylindrical velocity space (V_R, V_phi, V_z). The algorithm yields five clusters, labelled A through E, which the authors associate with known nearby substructures: the Helmi streams, M18-Cand10/MMH-1, Sequoia, Antaeus, and ED-2. Using DESI metallicities, they find that Clusters A, B, and E are chemically distinct from their local kinematic background, while Clusters C and D are not. The paper also estimates metallicity dispersions to tentatively infer progenitor types, associating Cluster A with a dwarf galaxy and Cluster E with a globular cluster. The central claim, stated in Section 5, is that the search confirms the presence of known substructures in DESI Y1 and thereby validates HDBSCAN* as a reliable tool for future blind substructure searches.","tokens_in":1566,"tokens_out":1677,"duration_ms":61937,"significance":"If the detections are genuine, the paper provides a useful demonstration that an unsupervised density-based clustering method can recover a set of known local halo substructures from DESI Y1 data, and it adds DESI metallicity measurements for these structures. The manuscript has several concrete strengths: the hyperparameter tuning is performed on 500 mock realisations, the clusters are validated with KS tests against local backgrounds, the perturbation stability is assessed over 200 iterations, and the full member catalogues and figure data are made public on Zenodo. The authors also include a candid limitations section. However, the validation is incomplete at a load-bearing point: the stability metric for four of the five clusters is low and is not calibrated against a smooth-halo null model, so the paper does not currently establish that Clusters B, C, D, and E are significant overdensities rather than marginal fluctuations. The claim that HDBSCAN* is validated for future discovery searches is therefore stronger than the evidence presented.","major_comments":[{"comment":"The stability analysis is not calibrated against a null model. Under the adopted recovery definition (a new cluster is a recovery if more than half of its members fall inside the original 99% ellipsoid), Clusters B, C, D, and E achieve IoM stability rates of only 23.5%, 27.0%, 25.0%, and 40.5%, respectively. The paper does not report what stability rate would be expected for a chance overdensity in a smooth halo under the same perturbations and selection function. The mock tuning in §2.4.2 provides completeness and purity for injected streams but never a false-positive rate for the adopted hyperparameters on a null-halo realisation. Without such a calibration, the numbers in Table 4 cannot distinguish genuine but noisy structures from marginal density fluctuations, which undermines the Section 5 claim that the detections confirm the presence of known substructures and validate HDBSCAN* for blind searches.","section":"§3.2, Table 4"},{"comment":"Clusters C and D are not chemically distinct from their local backgrounds, with KS p-values of 0.065 and 0.476, respectively. The authors acknowledge this in §3.1 but nevertheless include C and D among the five confirmed dynamic groups and use them for the Sequoia and Antaeus associations in §4.1.3. Because Clusters C and D also have low IoM stability (27% and 25% in Table 4), their reality currently rests entirely on the uncalibrated density clustering. Please either reclassify Clusters C and D as tentative candidates pending a null-halo calibration, or provide the quantitative false-positive test that establishes them as significant overdensities.","section":"§3.1, Figure 8"},{"comment":"The hyperparameter selection depends on mock priors whose representativeness is not demonstrated. The mock population assumes a composition of 35% thick disc, 50% GSE, and 15% other halo stars, with the GSE fraction deliberately overrepresented relative to the 15-25% literature range quoted by the authors, and with Gaussian velocity distributions. The adopted hyperparameters, (min_cluster_size, min_samples) = (20, 20) in IoM space and (10, 20) in velocity space, are then applied to the real data. No sensitivity analysis is shown for plausible variations in the mock composition, non-Gaussianity, or distance-error model, even though these choices determine all downstream cluster memberships and associations. A robustness test of the clustering output to these assumptions would materially strengthen the central validation claim.","section":"§2.4.1, §2.4.2"}],"minor_comments":[{"comment":"There is a numeric inconsistency in the halo subset size: the abstract states 138,661 stars, while §2.3 states that 138,861 stars remain in the halo subset. Please correct the inconsistent value.","section":"Abstract and §2.3"},{"comment":"The text refers to 'the bottom-left panel in Figure 4' when describing the large IoM spread of rejected Cluster E stars, but Figure 4 shows mock completeness and purity curves; the intended reference appears to be a panel in Figure 6 or Figure 11. Please fix this cross-reference.","section":"§4.1.4"},{"comment":"The parameter name 'min_sample_size' is a typo; HDBSCAN* uses min_samples, and the surrounding text in §2.4 consistently uses min_samples. Please correct the typo.","section":"§4.3.2"},{"comment":"The text states that Cluster E initially contains 44 stars, while Table 3 lists N_tot = 43 for Cluster E. This discrepancy should be reconciled.","section":"§4.1.4 and Table 3"}],"recommendation":"major_revision","confidential_remarks":"The paper is within scope for MNRAS and the data products will be useful to the Galactic archaeology community. The main obstacle is not the method itself but the overinterpretation of the stability and chemical validation: a smooth-halo false-positive test and a sensitivity analysis of the mock-based hyperparameter choice would make the reliability claim supportable. The authors' own limitations section shows they are aware of several of these issues, so I regard the requested changes as within the scope of a revision rather than grounds for rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Bottom line: this is a careful, honest application of HDBSCAN* to the DESI Y1 local halo sample. All five detected clusters match previously reported substructures, and the paper is upfront that nothing new was discovered. What it does add is a homogeneous set of DESI metallicities and clean member lists for the Helmi streams, M18-Cand10/MMH-1, Sequoia, Antaeus, and ED-2 within 5 kpc, plus a refreshingly candid limitations section.\n\nThe method description is thorough: 500 mock datasets for hyperparameter tuning, KS tests against background, and a 200-iteration perturbation analysis. The public member tables and Zenodo data are a real plus. For what it is — a recovery and chemodynamic characterization paper — it is solid.\n\nThe soft spots are real but not fatal. The stability metric in §3.2 is threshold-based and never calibrated against a null model. Four of five clusters have IoM stability between 23% and 40%, which sounds fragile; without a false-positive rate on a smooth-halo mock, you cannot tell whether that is normal for small genuine clusters under realistic distance errors or a sign of marginal density fluctuations. That matters because the paper's concluding claim is that HDBSCAN* is validated for future blind searches. Clusters C and D are not chemically distinct from background (KS p=0.065 and 0.476), Cluster E's globular-cluster classification depends on trimming four metal-rich members, and the Cluster B association relies partly on private correspondence. These are all addressable, and the external matching to known structures provides independent confirmation that the detections are not purely noise.\n\nWho is this for? Anyone working on nearby halo substructures, especially with DESI, will want the member tables and metallicity measurements. As a method-validation paper, it is not there yet.\n\nRecommendation: yes, send it to peer review. The data products and the recovery of known structures deserve referee time. The revision should add a null-model false-positive test and a calibrated stability statistic before claiming reliability for future blind searches.","headline":"A careful, honest recovery of five known halo substructures in DESI Y1, with useful DESI metallicities and public member tables, but the validation claim for future blind searches is only partially supported without a null-model calibration.","tokens_in":44524,"tokens_out":2025,"would_cite":true,"duration_ms":24142,"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":"Applying an automated density-based clustering search to 138,661 nearby stars in the DESI Milky Way Survey Year 1 data, this paper reports five kinematic groups that match the Helmi streams, M18-Cand10/MMH-1, Sequoia, Antaeus, and ED-2…","keywords":["Galactic halo","stellar substructures","DESI Milky Way Survey","HDBSCAN*","integrals of motion","halo chemodynamics","tidal debris","nearby stars"],"falsifier":"A direct test is to rerun the same clustering pipeline on the next DESI data release with improved parallaxes: if the five groups do not reappear as overdensities with similar orbits and higher membership, the claim that they are genuine known substructures would be falsified. A cheaper check is to lower min_cluster_size in integrals-of-motion space to 5 and switch to leaf selection; if the metallicity contrast of Clusters B, C, D, and E disappears or the groups merge into the background, the detections are hyperparameter artefacts.","tokens_in":43196,"feed_emoji":"🌌","tokens_out":7078,"duration_ms":65479,"temperature":0.7,"pith_summary":"The paper applies an unsupervised clustering algorithm, HDBSCAN*, to 138,661 nearby halo stars from the DESI Milky Way Survey Year 1 and claims it recovers five known stellar substructures: the Helmi streams, M18-Cand10/MMH-1, Sequoia, Antaeus, and ED-2. The detections are made independently in two spaces, one built from orbital integrals and one from Galactocentric cylindrical velocities. DESI metallicities show that three of the groups are chemically distinct from local halo stars with similar orbits, and metallicity dispersion points to a dwarf-galaxy origin for the Helmi streams and a globular-cluster origin for ED-2. If the result is correct, the search did not uncover new substructures, but it validates HDBSCAN* as a reliable automated tool for finding genuine ones in future data releases.","feed_headline":"No new streams: DESI search finds five known halo groups","feed_subtitle":"Unsupervised clustering of 138,661 DESI stars rediscovers five known halo streams, validating the method.","key_machinery":"The central machinery is HDBSCAN*, a density-based unsupervised clustering algorithm that finds over-dense groups without prior labels, run with the default 'eom' cluster-selection method. It is applied to a four-dimensional integrals-of-motion space (total energy, vertical angular momentum, and log-scaled radial and vertical actions) and separately to a three-dimensional Galactocentric cylindrical velocity space. The two most important hyperparameters, min_cluster_size and min_samples, were tuned on 500 mock halo datasets containing 15 injected streams, leading to adopted values of (20,20) in integrals-of-motion space and (10,20) in velocity space. Overlapping detections in the two spaces are merged into five clusters, then validated with DESI metallicities, perturbation-based stability rates, and comparisons with literature member lists.","core_discovery":"The central claim is that a density-based unsupervised clustering search of the DESI Year 1 halo subset does not need prior labels to find substructures: it independently identifies five kinematically coherent groups within 5 kpc, and every one matches a previously reported substructure. The authors state explicitly that the search did not lead to the discovery of new substructures, but that the five detections confirm the presence of known substructures and validate the reliability of using HDBSCAN* for such searches. Using DESI metallicities, they show that the Helmi streams, M18-Cand10/MMH-1, and ED-2 are chemically distinct from the local halo, while Sequoia and Antaeus are not clearly distinct by a Kolmogorov–Smirnov test. The paper also uses metallicity dispersion to tentatively associate the Helmi streams with a dwarf-galaxy progenitor and ED-2 with a globular-cluster progenitor, leaving the origins of the other three groups ambiguous.","pith_inferences":["Editorial inference: The failure to detect GSE, the thick disc, and most globular clusters is likely a consequence of the tuned hyperparameters and the distance-error cut, not a limitation of DESI data; a two-stage search using leaf clustering or a smaller min_cluster_size could recover cold streams that are currently missed.","Editorial inference: The low stability of Clusters B, C, and D in integrals-of-motion space suggests that their associations with M18-Cand10/MMH-1, Sequoia, and Antaeus could shift if improved astrometry moves a handful of member stars, and chemical tagging with elements such as C, Mg, and Ca would provide a sharper test.","Editorial inference: If Sequoia and Antaeus are fragments of a single massive accretion event rather than independent mergers, their overlapping chemodynamic spaces imply that counting overdensities alone may overestimate the number of distinct accretion events in the inner halo."],"forward_implications":["The five recovered groups show that the DESI Year 1 halo subset contains the Helmi streams, M18-Cand10/MMH-1, Sequoia, Antaeus, and ED-2 within 5 kpc.","DESI metallicities provide a clean chemical confirmation for at least three of these groups, and metallicity dispersion associates the Helmi streams with a dwarf-galaxy progenitor and ED-2 with a globular-cluster progenitor.","Because HDBSCAN* recovered these groups without prior labels, the same automated pipeline can be applied to later DESI releases to search for new substructures with reduced selection bias.","Most clusters have low stability under 200 perturbations of the phase-space measurements in integrals-of-motion space, with only Cluster A reaching 100 percent stability, so future detections will require better distances and larger samples to be secure.","HDBSCAN* with these settings does not detect GSE, the thick disc, or known globular clusters as separate structures, indicating that the method is tuned for compact overdensities rather than broad or diffuse components."],"supporting_citations":[{"why":"Provides the original discovery of the Helmi streams and the integrals-of-motion approach that Cluster A is compared against.","marker":"Helmi et al. (1999)"},{"why":"Supplies the core member list of Helmi streams used for the chemodynamic comparison with Cluster A.","marker":"Koppelman et al. (2019a)"},{"why":"Supplies M18-Cand10 member kinematics and metallicity used to associate Cluster B.","marker":"Myeong et al. (2018a)"},{"why":"Provides the MMH-1 velocity feature used as an alternate association for Cluster B.","marker":"Mikkola et al. (2023)"},{"why":"Supplies Antaeus action properties and metallicity used to associate Cluster D.","marker":"Oria et al. (2022)"},{"why":"Supplies the ED-2 member list and orbital parameters used to associate Cluster E.","marker":"Dodd et al. (2023)"},{"why":"Provides high-resolution spectroscopy establishing ED-2's globular-cluster origin, used for Cluster E's progenitor interpretation.","marker":"Balbinot et al. (2024)"},{"why":"Supplies the HDBSCAN* algorithm implementation used for all clustering in the paper.","marker":"McInnes et al. (2017)"},{"why":"Supplies the velocity distributions for thick disc, GSE, and inner halo stars used to construct the mock datasets for hyperparameter tuning.","marker":"Belokurov et al. (2020)"}],"fun_headline_variants":["DESI halo search rediscovers five known groups, no new ones","HDBSCAN* on DESI stars: five known substructures, zero new","No new halo streams; DESI clustering confirms five known ones","Five previously seen halo structures found in DESI data","DESI Y1 halo clumps: all are familiar, method passes test"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the mock datasets used to tune HDBSCAN* represent the real composition of the local halo—35 percent thick disc, 50 percent GSE-like stars, and 15 percent other halo, with Gaussian velocity distributions—so that the chosen hyperparameters are calibrated for the true data; if that composition is wrong, the detected clusters could be tuning artefacts rather than real substructures.","fun_headline_variants_meta":{"raw":{"variants":["DESI halo search rediscovers five known groups, no new ones","HDBSCAN* on DESI stars: five known substructures, zero new","No new halo streams; DESI clustering confirms five known ones","Five previously seen halo structures found in DESI data","DESI Y1 halo clumps: all are familiar, method passes test"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000626,"raw_usage":{"total_tokens":2955,"prompt_tokens":1059,"completion_tokens":1896,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":675,"completion_tokens_details":{"reasoning_tokens":1802}},"tokens_in":675,"tokens_out":1896,"duration_ms":13905,"temperature":1.0,"reasoning_tokens":1802,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T05:32:22.300595+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A direct test is to rerun the same clustering pipeline on the next DESI data release with improved parallaxes: if the five groups do not reappear as overdensities with similar orbits and higher membership, the claim that they are genuine known substructures would be falsified. A cheaper check is to lower min_cluster_size in integrals-of-motion space to 5 and switch to leaf selection; if the metallicity contrast of Clusters B, C, D, and E disappears or the groups merge into the background, the detections are hyperparameter artefacts.","supporting_citations":[],"review_version":1}