{"id":"b98c4997-f4bc-4c57-8f41-c1f28090b205","arxiv_id":"2504.20155","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"The 2017 KMTNet subprime-field search adds four unambiguous planets, completing a homogenous 112-planet microlensing sample that confirms the 'sub-Saturn desert' in planet-host mass ratios.","lead":"Astronomers used a network of three telescopes to find four new planets by watching their host stars briefly brighten due to microlensing, completing a four-year search catalog of 112 planets. The full catalog shows an unexpected shortage of planets with masses around that of sub-Saturn, which challenges and informs planet formation models.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The sub-Saturn desert 'confirmation' rests on raw 2016/2017 counts, not the detection-efficiency correction used for 2018/2019, so the abstract overstates the evidence.","rationale":"The paper's central claims are the completeness of the 112-planet 2016-2019 KMTNet AnomalyFinder sample and the confirmation of the sub-Saturn desert by the 2016/2017 data. The completeness claim is plausible but rests partly on a manual operator step that is not independently validated. I find the more immediately load-bearing issue to be the desert confirmation: the paper compares raw, uncorrected mass-ratio distributions for 2016/2017 with an efficiency-corrected 2018/2019 result and calls that a confirmation. Since the paper itself states that no mass-ratio-function analysis is attempted for this sample, the raw plateau in Figure 15 is subject to the same selection effects that the 2018/2019 analysis needed to remove. The four new planet analyses appear careful and standard, with 1L2S checks, caustic-geometry exploration, and Bayesian physical-parameter estimates, so I do not see a fatal flaw. The concern is addressable by applying the efficiency correction already developed in the series to the 2016/2017 fields, which would settle whether the desert confirmation survives. This matches part of the reader's weakest assumption, though I focus on the efficiency correction rather than the manual completeness audit.","tokens_in":28231,"tokens_out":9975,"duration_ms":102210,"concrete_test":"Recompute the 2016-2017 mass-ratio function with the same injection/recovery efficiency correction used by Zang et al. (2025) for 2018-2019: inject synthetic 2L1S anomalies into the actual 2016 and 2017 KMTNet cadences across all fields, run them through AnomalyFinder and the operator triage step, and divide the observed q distribution by the recovery fraction as a function of (log q, s, t_E, source magnitude). If the [-3.6,-3.0] deficit persists at comparable significance, the abstract's word 'confirmed' is justified; if it disappears or shrinks materially, the desert claim for 2016/2017 must be downgraded to 'consistent' or 'not independently confirmed.'","verdict_should_be":"UNCHANGED","load_bearing_attack":"The abstract's claim that the sub-Saturn desert is 'confirmed' by the 2016 and 2017 KMTNet samples is the most load-bearing assertion and is not supported by the analysis presented. The 2018/2019 desert was established only after correcting for KMTNet/AnomalyFinder detection efficiency (Zang et al. 2025). For 2016/2017 this paper provides only raw counts: Figure 15 plots the cumulative mass-ratio distribution of detected planets, and Section 5 explicitly says it does not attempt to study the mass-ratio function. A raw deficit in log q = [-3.6,-3.0] can be produced by selection effects (including the manual triage of 3315 AnomalyFinder candidates by a single operator and the differing sensitivity of by-eye vs AnomalyFinder discovery) just as easily as by an intrinsic desert. Comparing this raw distribution with the heterogeneously selected pre-KMTNet sample is not a controlled test. Thus the appropriate conclusion is that the 2016/2017 data are consistent with the desert, not that they confirm it; the confirmation would require an efficiency-corrected analysis of the 2016/2017 fields.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper completes the KMTNet AnomalyFinder planetary sample for the 2017 subprime fields by analyzing four unambiguous planets (KMT-2017-BLG-0849, KMT-2017-BLG-1057, OGLE-2017-BLG-0364, KMT-2017-BLG-2331) and one candidate (KMT-2017-BLG-0958). The light-curve analysis follows the series standard: 2L1S grid searches with VBBinaryLensing, checks of close/wide topologies, 1L2S alternatives, and higher-order effects, followed by Bayesian estimates of host and planet masses. The authors combine these with previously published and forthcoming events to form a sample of 112 unambiguous planets from 2016–2019, which they state nearly triples the microlensing planetary sample, and they claim that the \\\"sub-Saturn desert\\\" (log q = [-3.6, -3.0]) found in the 2018–2019 KMTNet samples is confirmed by the 2016–2017 samples.","tokens_in":28438,"tokens_out":6922,"duration_ms":61193,"significance":"If the completeness claims hold, the 112-planet AnomalyFinder sample is the first large homogeneous microlensing sample spanning four seasons, and it will be a valuable resource for mass-ratio function studies. The individual event analyses are careful, including explicit treatment of the Cannae/von Schlieffen degeneracy for KMT-2017-BLG-0849 and the close/wide degeneracy for OGLE-2017-BLG-0364, with the full Δχ² values reported so readers can apply alternative thresholds. The Bayesian physical-parameter estimates use external Galactic-model priors and lens-flux upper limits, which is appropriate. The primary weakness is that the paper's headline claim—that the sub-Saturn desert is \\\"confirmed\\\" by the 2016–2017 data—is based on raw counts rather than the efficiency-corrected analysis used to establish the desert in 2018–2019.","major_comments":[{"comment":"The claim that the sub-Saturn desert is \\\"confirmed\\\" by the 2016–2017 KMTNet samples is not supported by the analysis as presented. The 2018–2019 desert was established only after correcting for KMTNet/AnomalyFinder detection efficiency (Zang et al. 2025), whereas Section 5 for 2016–2017 uses only raw counts in Figure 15 and explicitly states that the authors \\\"do not attempt to study the mass-ratio function here.\\\" A raw deficit in log q = [-3.6, -3.0] could be produced by selection effects, including the manual triage of 3315 AnomalyFinder candidates by a single operator and the different sensitivities of by-eye versus AnomalyFinder discovery. Comparing this raw distribution with the heterogeneous pre-KMTNet sample is not a controlled test. Please rephrase the abstract and Section 5 to say that the 2016–2017 data are consistent with the desert, or provide an efficiency-corrected analysis of the 2016–2017 fields.","section":"Abstract and Section 5, Figure 15"},{"comment":"The paper adopts a degeneracy criterion (exclude models with Δχ² > 20) that is looser than the series standard (Δχ² > 10), but Table 8 uses a different rule, listing only models with Δχ² < 10 relative to the best fit. For OGLE-2017-BLG-0364, the Close Inner model is disfavored by only Δχ² = 18.9 and is retained as a viable solution in the light-curve and Bayesian analyses (Tables 4 and 7), yet Table 8 reports only log q = -3.341. If Close Inner is correct, log q ≈ -2.855, a materially different planet that lies just outside the desert interval. The sample definition and the \\\"unambiguous\\\" classification need a single, clearly stated threshold, and the table should either list the degenerate solutions or explain why they are omitted from the final sample.","section":"Section 3.1 and Table 8"},{"comment":"The completeness of the 2016–2017 sample—and hence the 112-planet \\\"complete sample\\\" claim and the raw mass-ratio distribution in Figure 15—rests on a manual step: \\\"the operator (W. Zang) identified 133 anomalous events\\\" from 3315 AnomalyFinder candidates. This operator-dependent triage is not validated, for example by an independent search or a reproducibility measure, and it is load-bearing for the completeness claim. Please quantify the reproducibility of this step or state clearly as a limitation that subtle anomalies may have been missed, which would affect the raw mass-ratio distribution and the desert interpretation.","section":"Section 1 and Section 5"}],"minor_comments":[{"comment":"The typo \\\"AnomlyFinder\\\" appears multiple times (Section 5, Table 8 note, Figure 15 labels) and should be corrected to \\\"AnomalyFinder.\\\"","section":"Section 5 and Table 8 note"},{"comment":"\\\"From the 2017 KNTNet subprime data\\\" should be \\\"2017 KMTNet subprime data.\\\"","section":"Section 1"},{"comment":"\\\"A PLPS fit by excluding the anomaly\\\" should read \\\"A PSPL fit.\\\"","section":"Section 3.4"},{"comment":"The sentence \\\"We expect that ρ and q are the lower limits because the caustic crossing time is ≤ 2tEρ for a larger source\\\" is unclear; consider rephrasing to state that the heuristic estimates are lower limits because finite-source effects can be larger than inferred from the caustic-crossing duration.","section":"Section 3.2"},{"comment":"The sentence listing the seasonal distribution (23, 30, 35, 24) states it is consistent with Poisson variations; adding a brief significance estimate (e.g., the standard deviation of a Poisson mean of 28 is about 5.3) would make this quantitative.","section":"Section 5"}],"recommendation":"major_revision","confidential_remarks":"The paper is a solid addition to the AnomalyFinder series, but the desert-confirmation claim in the abstract is stronger than the raw-count analysis supports. The editor may wish to ensure that the final version either includes an efficiency-corrected analysis for 2016–2017 or tempers the claim to \\\"consistent with\\\" the desert. The loose degeneracy threshold and the inconsistency between Section 3.1 and Table 8 also need reconciliation, as they affect the reproducibility of the sample definition."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"What matters here: the paper completes the 2016-2019 KMTNet AnomalyFinder sample, adding four unambiguous planets and one candidate, and assembles a 112-planet catalog that nearly triples the published microlensing sample. The individual light-curve modeling is the best part. Each event gets close/wide, 2L1S vs 1L2S, multiple wide local minima, and 'hotter' MCMC checks; OGLE-2017-BLG-0364 in particular is handled honestly with four competing models, and the paper gives Delta-chi-squared values so the reader can apply a stricter criterion. Bayesian host and planet masses use standard Galactic priors and lens-flux upper limits. This is a workmanlike capstone, and the catalog will be used.\n\nThe main soft spot is the sub-Saturn desert 'confirmation.' The 2018/2019 desert was established only after correcting for detection efficiency. Here the 2016 and 2017 samples are presented as raw counts: Figure 15 is cumulative counts, and Section 5 explicitly says the mass-ratio function is not studied. A raw deficit in log q = [-3.6,-3.0] can arise from selection effects, including the manual triage of 3315 AnomalyFinder candidates by one operator and the different sensitivity of by-eye versus algorithm discovery. So the appropriate claim is that 2016/2017 are consistent with the desert, not that they confirm it. That is a matter of one word in the abstract and one sentence in the discussion, but it matters because the paper's significance partly rides on it.\n\nSecond, the loose Delta-chi-squared greater than 20 degeneracy threshold: for OGLE-2017-BLG-0364, the Close Inner model is only Delta-chi-squared = 18.9 from the best and has log q = -2.855, outside the desert. The planet's inclusion with a model inside the desert is thus somewhat uncertain. The authors are transparent about choosing a loose criterion, but the desert-edge statistics inherit that sensitivity.\n\nThird, completeness rests on the manual step in Section 1. The operator found 133 anomalous events from 3315 candidates, with no independent validation and no released code. This is not fatal because the series has used the same procedure, but it means 'complete' is not machine-checked.\n\nThis is for microlensing and exoplanet demographics readers. I would send it to a serious referee. The new planets are real, the catalog is valuable, and the flaws are fixable. I would ask for a softened desert claim or an efficiency-corrected 2016/2017 check, a note on how the threshold choice affects sample membership, and ideally release of the anomaly list.","headline":"Four new planets are real and the completed 112-planet catalog is valuable, but the sub-Saturn desert 'confirmation' rests on raw counts and should be softened or efficiency-corrected.","tokens_in":29334,"tokens_out":2742,"would_cite":true,"duration_ms":27538,"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 2017 subprime-field analysis completes a 112-planet KMTNet sample that confirms the 'sub-Saturn desert' in planetary mass ratios.","keywords":["gravitational microlensing","exoplanet detection","KMTNet","AnomalyFinder","planetary mass ratio","sub-Saturn desert","Galactic bulge","light-curve modeling"],"falsifier":"Re-run the complete 2017 subprime search with the same data but with two independent reviewers classifying all 3,315 candidates, and also inject synthetic planet signals into the light curves to measure what fraction are recovered. If the final planet count or the number of planets with $\\log q$ between $-3.6$ and $-3.0$ changes by more than the reported uncertainties, the claimed completeness and the confirmed desert would be contradicted.","tokens_in":27935,"feed_emoji":"🪐","tokens_out":14297,"duration_ms":126493,"temperature":0.7,"pith_summary":"This paper completes the final piece of a systematic census of planets found by gravitational microlensing in the first four years of the KMTNet survey: the low-cadence 'subprime' fields of 2017. It reports four new unambiguous planets—KMT-2017-BLG-0849, KMT-2017-BLG-1057, OGLE-2017-BLG-0364, and KMT-2017-BLG-2331—with planet-host mass ratios from $1.0\\times10^{-4}$ to $1.3\\times10^{-3}$, plus one candidate, KMT-2017-BLG-0958, whose light curve cannot be conclusively separated from a binary-source interpretation. Once merged with previously reported planets, the 2016-2019 AnomalyFinder sample contains 112 unambiguous planets, about three times the number of microlensing planets known before KMTNet, and nearly half of them come from the low-cadence subprime fields. The paper's key statistical claim is that this complete sample confirms the 'sub-Saturn desert,' a deficit of planets with mass ratios $\\log q$ between $-3.6$ and $-3.0$, previously seen in 2018-2019 and now supported by the complete 2016-2017 seasons.","feed_headline":"112 planets confirm the sub-Saturn desert","feed_subtitle":"Complete 2016-2019 KMTNet census nearly triples the microlensing sample and shows the desert is real.","key_machinery":"The machine that carries the argument is the AnomalyFinder pipeline, applied uniformly to every KMTNet event: each light curve is fitted with a point-source point-lens (PSPL) model, the residuals are scanned for anomalies, and the anomalous events are re-fitted with 2L1S binary-lens models, in which the planet-host mass ratio $q$ and projected separation $s$ are the parameters of interest. Completeness comes from the pipeline, not from any single event: the same search and the same model-selection rules have now been applied to all four seasons and to both prime and subprime fields. Within individual events, the load-bearing analysis tools are the 'hotter' Markov-chain Monte Carlo search that maps out competing caustic-crossing topologies, the 1L2S (binary-source) check that separates genuine planets from two-star sources, and a deliberately permissive $\\Delta\\chi^2 > 20$ rule for discarding alternative models, with all surviving models reported so that readers can apply their own threshold.","core_discovery":"The central claim, stated the way the authors would state it, is that the first four years of KMTNet data now hold a complete, homogeneously searched planetary sample, and that its mass-ratio distribution is a fair census of the planets the survey could detect. After the 2017 subprime analysis reported here, the combined 2016-2019 AnomalyFinder sample contains 112 unambiguous planets, with a seasonal distribution $(23,30,35,24)$ consistent with Poisson fluctuations; of these, 37 were first found by the systematic search and 75 were previously known from by-eye searches and then recovered. The paper finds that the 2017 subprime sample alone contains 15 unambiguous planets, of which 5 have $\\log q < -3.0$. The mass-ratio distribution of the full sample shows a plateau, or desert, between $\\log q = -3.6$ and $-3.0$, and the paper argues this now confirms the desert reported from the 2018 and 2019 seasons, making it a genuine feature of planets around M and K dwarfs rather than an artifact of incomplete searching. The new planets add mass ratios $(1.0,1.2,4.6,13)\\times10^{-4}$, corresponding to median planetary masses of $6.4, 24, 76$, and $171\\,M_\\oplus$ around $0.2-0.6\\,M_\\odot$ hosts.","pith_inferences":["If the desert is a real feature of the planet population, it should reappear in the 2020-2022 KMTNet seasons and in future bulge surveys; a filled-in desert there would instead point to a time-varying selection effect.","The paper's permissive $\\Delta\\chi^2 > 20$ threshold means borderline events (for example the candidate KMT-2017-BLG-0958 and the 'Wide B'/'Wide C' models of OGLE-2017-BLG-0364) sit close to the sample boundary, so a uniform re-analysis across all seasons with a single stricter threshold would test how much the desert's edges move.","The desert spans roughly the planet-mass range of the 'planet gap' or 'Neptune desert' seen in radial-velocity and transit demographics around sun-like stars; if the two deficits share a formation origin, microlensing offers the only current way to test that origin at wide separations and around M-dwarf hosts.","The unresolved 1L2S degeneracy in the candidate event hints that small numbers of 'possible' planets may exist in all seasons, so the true planet count could be slightly higher than 112."],"forward_implications":["The 112-planet complete sample provides a homogeneous statistical basis for planet demographics at separations beyond the snow line, the regime where microlensing is uniquely sensitive.","The confirmed sub-Saturn desert at $\\log q \\in [-3.6,-3.0]$ becomes a feature that any formation model—most directly the core-accretion runaway-growth scenario—must reproduce for low-mass host stars.","Because 37 of the 112 planets were found only by the systematic search, future microlensing planet samples that rely on by-eye detection will be incomplete at low mass ratios, motivating automated searches in ongoing and planned surveys.","The 2016-2019 KMTNet sample reaches mass ratios an order of magnitude smaller than pre-KMTNet samples and increases the number of $\\log q < -4$ planets five-fold, sharpening measurements of the low-ratio slope of the mass-ratio function."],"supporting_citations":[{"why":"Introduces the AnomalyFinder algorithm and the goal of a homogeneous KMTNet planetary sample, providing the search method this paper applies to the 2017 subprime fields.","marker":"Zang et al. 2021a"},{"why":"Forms the efficiency-corrected 2018-2019 statistical sample and identifies the sub-Saturn desert that this paper's complete 2016-2017 sample is said to confirm.","marker":"Zang et al. 2025"},{"why":"Reported the first hint of the desert in 2016-2017 by-eye planets but attributed it to publication bias; the current complete sample overturns that interpretation.","marker":"Yang et al. 2020"},{"why":"Defines the point-source point-lens baseline model whose residuals AnomalyFinder searches for anomalies.","marker":"Paczyński 1986"},{"why":"Describes the EventFinder algorithm and the KMTNet field and cadence layout that separates prime from subprime fields, the completeness units used here.","marker":"Kim et al. 2018"},{"why":"Establishes the 2L1S/1L2S degeneracy used to classify events as unambiguous planets versus candidates.","marker":"Gaudi 1998"},{"why":"Published three of the 2017 subprime unambiguous planets and set the light-curve analysis conventions this paper follows.","marker":"Zang et al. 2023"},{"why":"Completed the 2016 subprime-field sample, the adjacent season whose methods and criteria this paper inherits.","marker":"Shin et al. 2024"},{"why":"Completed the 2017 prime-field sample, making the 2017 season fully covered once the present subprime analysis is added.","marker":"Ryu et al. 2024"},{"why":"Completed the 2019 subprime-field sample and demonstrated the complete-sample analysis that the 2017 subprime work extends.","marker":"Jung et al. 2023"}],"fun_headline_variants":["112 planets confirm the sub-Saturn desert","Complete KMTNet 2016-2019 sample: 112 planets, desert confirmed","2017 subprime fields complete KMTNet census at 112 planets","4-year KMTNet search: 112 planets, sub-Saturn desert confirmed"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the 2016-2017 data really are complete: every planet-like anomaly among the 3,315 candidate signals was caught by one person's manual review, and the deliberately loose rule for discarding alternative explanations did not let any impostor into the final 112-planet list.","fun_headline_variants_meta":{"raw":{"variants":["112 planets confirm the sub-Saturn desert","Complete KMTNet 2016-2019 sample: 112 planets, desert confirmed","2017 subprime fields complete KMTNet census at 112 planets","4-year KMTNet search: 112 planets, sub-Saturn desert confirmed"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000738,"raw_usage":{"total_tokens":3409,"prompt_tokens":1169,"completion_tokens":2240,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":785,"completion_tokens_details":{"reasoning_tokens":2159}},"tokens_in":785,"tokens_out":2240,"duration_ms":16519,"temperature":1.0,"reasoning_tokens":2159,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T05:36:29.195874+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the complete 2017 subprime search with the same data but with two independent reviewers classifying all 3,315 candidates, and also inject synthetic planet signals into the light curves to measure what fraction are recovered. If the final planet count or the number of planets with $\\log q$ between $-3.6$ and $-3.0$ changes by more than the reported uncertainties, the claimed completeness and the confirmed desert would be contradicted.","supporting_citations":[{"cited_title":"K., Yang , H., et al","cited_arxiv_id":null,"evidence_quote":"Published three of the 2017 subprime unambiguous planets and set the light-curve analysis conventions this paper follows."},{"cited_title":"K., Zang , W., Wang , H., et al","cited_arxiv_id":null,"evidence_quote":"Completed the 2019 subprime-field sample and demonstrated the complete-sample analysis that the 2017 subprime work extends."}],"review_version":1}