{"id":"b37997b4-13f5-470c-91fe-5b8a597331a1","arxiv_id":"2504.15916","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"In 5 of 6 Gaussian feature tests, the K2-18b MIRI/LRS spectrum prefers a flat line over spectral features, with only weak evidence (ln(B)=1.21) for features at fixed wavelengths.","lead":"A reanalysis of JWST's MIRI/LRS spectrum of the sub-Neptune K2-18b finds that a simple flat line fits the data as well as Gaussian-shaped spectral features. The study challenges the recent 3.4-sigma claim of dimethyl sulphide or disulphide detections and suggests the atmosphere's composition remains unproven.","discovery_kind":"replication","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Gaussian-only feature search undercuts the broad claim of no spectral features.","rationale":"The paper is careful in its reduced chi-squared and evidence calculations for the Gaussian models, and the finding that a flat line is acceptable is an honest result. However, the central claim is broader than the analysis supports. The model family (single or double Gaussian) is a proxy for the molecular features in question, and the paper's assertion of 'Gaussian-like' shapes is an unverified assumption that directly feeds the null conclusion. This is a load-bearing gap because the original detection claim rests on a different model family. The proposed test—computing the evidence with actual molecular profiles—would settle whether the flat-line preference is robust to the shape assumption. The reader identified the data uncertainties and unstated priors as the weakest points; I agree on the priors but consider the model-family mismatch more significant. Therefore, conditional acceptance is appropriate: the paper should either restrict its conclusion to Gaussian features or include a full-model comparison.","tokens_in":5109,"tokens_out":8875,"duration_ms":84731,"concrete_test":"Re-compute the Bayes factor between a flat line and a physically motivated DMS/DMDS model (e.g., the best-fit retrieval from Madhusudhan et al. 2025, or a simplified band model using the actual absorption cross-sections) on the same JexoRes binned data, with shared priors and the same MultiNest settings. If this full model yields ln(B) >= 2.5 relative to flat, the Gaussian-only analysis misses real spectral features and the conclusion must be narrowed; if ln(B) < 1, the concern is resolved.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The paper's central conclusion (Section 4: 'no strong evidence for detected spectral features') is inferred solely from tests in which the alternative to a flat line is one or two Gaussians (Eq. 1). The only positive result is the fixed-wavelength two-Gaussian model with ln(B)=1.21, which is weak evidence. However, Madhusudhan et al. (2025) claimed detection on the basis of full atmospheric retrievals with molecular opacities, not Gaussian profiles. The paper assumes DMS/DMDS features are 'approximately Gaussian' (Section 2), but that assumption is not tested and is not generally true for molecular bands; individual lines may be Gaussian, but a band envelope can have non-Gaussian structure (e.g., P/R branch heads). If the actual absorption shape deviates from the assumed Gaussian form, the evidence comparison is biased against the feature: the extra flexibility of the Gaussian model is penalized by the Occam factor even if a physically motivated model would fit well. Consequently, the flat-line preference in 5 of 6 tests does not rule out the presence of molecular features; it only rules out Gaussian-shaped features. The conclusion overgeneralizes from the Gaussian model family to 'spectral features' broadly. Additionally, the Bayesian evidence values are not reproducible because the prior ranges for the Gaussian amplitude and width are not stated (Section 2), making the Jeffreys-scale classification untestable.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This research note reanalyzes the JWST MIRI/LRS transmission spectrum of K2-18b published by Madhusudhan et al. (2025). The author fits a flat line and six Gaussian-feature models (positive, negative, and combined, with free or fixed central wavelengths) to the JexoRes binned data and compares Bayesian evidences computed with MultiNest. Five of the six Gaussian models yield ln(B) < 1 relative to the flat line, and the most favorable model, two Gaussians fixed at ~7 and ~8.8 um, gives ln(B) = 1.21, which is classified as weak evidence on the Jeffreys scale. The paper concludes that there is no strong statistical evidence for spectral features and argues that the original 3.4-sigma detection claim is not supported by the MIRI/LRS data alone.","tokens_in":5341,"tokens_out":5137,"duration_ms":48956,"significance":"If the conclusion holds, this is a useful independent check on a high-profile biosignature claim: it would show that the MIRI/LRS data do not, by themselves, strongly endorse the DMS/DMDS detection claimed by Madhusudhan et al. (2025). The paper is transparent about the data source, reports chi-squared values for every model, and avoids the circularity of reusing fitted parameters as hypotheses. The fixed-wavelength tests are a legitimate targeted hypothesis, motivated by the original paper's own figures. However, the significance is limited by the narrow model family considered: all tests assume Gaussian-shaped features, and no injection-recovery or sensitivity test is performed. The reported evidence values also depend on unstated prior ranges and sampler settings. The result is therefore best read as 'no evidence for Gaussian-shaped features' rather than as a general statement about the absence of spectral features.","major_comments":[{"comment":"The central conclusion in Section 4 ('no strong evidence for detected spectral features') is inferred from tests in which the only non-flat model is one or two Gaussians. The manuscript asserts that DMS/DMDS spectral shapes are 'approximately Gaussian' by citing Figure 3 of Madhusudhan et al. (2025), but this assumption is not independently tested. Molecular band envelopes, especially with P/R branch structure, can deviate substantially from a Gaussian, and the Bayesian model comparison will then penalize the Gaussian model through the Occam factor even when a physically motivated molecular model fits the data well. Because the title, abstract, and conclusion make a broad claim about 'spectral features,' this model-family restriction is load-bearing. I recommend either explicitly restricting the conclusion to Gaussian-shaped features or adding an injection-recovery test that includes both Gaussian and non-Gaussian (e.g., molecular opacity) synthetic signals.","section":"Sec. 2, Eq. (1)"},{"comment":"The reported Bayesian evidences are not reproducible. No prior ranges are stated for the Gaussian amplitude A, width sigma_m, or vertical offset c, and no MultiNest settings are provided (e.g., number of live points, evidence tolerance, sampling efficiency). Since the Bayes factors and hence the Jeffreys-scale classifications in Section 3 depend directly on the prior volumes, the 'weak evidence' statement and the quantitative ln(B) values cannot be checked by readers. Please provide the full prior specification and sampler configuration, and ideally release the code or configuration files.","section":"Sec. 2"},{"comment":"The analysis contains no sensitivity or injection-recovery test. A finding of 'no evidence' is ambiguous: it can mean that no spectral feature is present, or that the data are too noisy, or that the adopted error bars are too large to permit detection of a feature of the expected amplitude. For instance, injecting a Gaussian with amplitude comparable to the claimed ~7 um DMS feature into the JexoRes binned data and checking whether ln(B) exceeds the weak-evidence threshold would calibrate the test's power. Without such a test, the flat-line preference cannot be interpreted as a strong upper limit on spectral features, especially since the analysis takes the JexoRes error bars at face value; underestimated errors would also produce the same qualitative pattern.","section":"Sec. 3"}],"minor_comments":[{"comment":"'Jeffrey's scale' should be 'Jeffreys scale' (and similarly in the abstract and conclusions).","section":"Sec. 3"},{"comment":"The displayed Gaussian formula contains a mismatched parenthesis: the argument of the exponential appears as '(lambda - mu)^2) / 2 sigma_m^2)' with an extra closing parenthesis. Please correct the equation formatting.","section":"Eq. (1)"},{"comment":"The figure would be easier to compare if the six panels were labeled (a)-(f) and the y-axis label were placed once rather than repeated on every panel, especially since the models are discussed as a suite in the text.","section":"Fig. 1"},{"comment":"The phrase 'Bayes factor = 3.35' would be clearer as 'Bayes factor of e^{1.21} = 3.35' or 'Bayes factor = 3.35 (since ln B = 1.21)' to avoid any confusion about logarithms.","section":"Sec. 3"},{"comment":"The repeated style 'N. Madhusudhan et al. 2025' is slightly awkward; 'Madhusudhan et al. (2025)' is more standard and avoids the author-initial prefix in prose.","section":"Abstract and Sec. 2"}],"recommendation":"major_revision","confidential_remarks":"This is a simple, readable research note with a clear negative result. The statistical core is internally consistent, and the lack of circularity is a genuine strength: the Gaussian tests are independent hypotheses, and the fixed-wavelength models are targeted rather than recycled from the fit. The main weakness is the mismatch between the broad conclusion ('no strong evidence for spectral features') and the narrow Gaussian-only model family; this is fixable by rewording and by adding an injection-recovery test. The missing prior ranges and sampler settings are also fixable. I do not think rejection is warranted, but the manuscript needs a substantive revision before I would recommend acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe short version: this research note applies the standard Gaussian feature search (May et al., JWST ERS) to the freshly published MIRI/LRS transmission spectrum of K2-18b and finds that in 5 of 6 model tests, a flat line is preferred; only the model with Gaussians fixed at 7 and 8.8 µm gives weak evidence (ln B = 1.21). That is a direct and useful challenge to the 3.4-σ feature claim in Madhusudhan et al. (2025). The analysis is transparent: it uses nested model comparison, reports χ² and ln(evidence), and makes a fair point that the original 'canonical' model is not a superset of the flat line, so prior ranges matter. I also think the note's observation about the significance drop between the 'canonical' and 'maximal' models is worth flagging.\n\nThe main soft spot is the interpretive leap. The title asks whether there are spectral features, and the abstract answers 'no strong statistical evidence for spectral features.' But the tests only compare Gaussian shapes. Molecular bands are not necessarily Gaussian; a P/R branch could be non-Gaussian and still be a real feature. So the paper can only claim no evidence for Gaussian-shaped features, not no spectral features. The original detection claim came from a full retrieval, and this note does not directly test that model family. The stress-test note makes this point, and I think it lands.\n\nThere are also reproducibility gaps: no prior ranges or MultiNest settings are given, and there is no sensitivity test showing the method would recover a known injected feature. These are easy to fix in a revised version. The fixed-wavelength test is post hoc but explicitly motivated by the earlier figure, so that's acceptable.\n\nOverall, this is a worthwhile, if limited, contribution. It won't settle K2-18b, but it should push the conversation toward more rigorous null tests. The prose is clear and the logic is sound within its stated framework. I would send it to peer review with comments asking for sensitivity tests and a more careful framing of the conclusion. Who's it for? Anyone working on exoplanet transmission spectroscopy or the K2-18b biosignature debate.\n\nRecommendation: accept for peer review, expect light-to-moderate revision.","headline":"A concise, timely Gaussian-feature reanalysis finds the flat line acceptable for K2-18b's MIRI/LRS spectrum, but the conclusion overreaches from 'no Gaussian features' to 'no spectral features.'","tokens_in":5872,"tokens_out":2870,"would_cite":true,"duration_ms":26506,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper re-examines the K2-18b MIRI/LRS spectrum and finds no strong statistical evidence for the spectral features behind the reported 3.4-sigma detection.","keywords":["K2-18b","MIRI/LRS transmission spectrum","Gaussian spectral features","Bayesian model comparison","dimethyl sulfide (DMS)","flat-line spectrum","JWST exoplanet atmospheres"],"falsifier":"A re-reduction of the raw MIRI/LRS time series that yields materially different transit depths near 7 and 8.8 microns, or an injection-recovery test showing that the method fails to recover a simulated feature at the claimed amplitude, would overturn or confirm the flat-line conclusion.","tokens_in":4892,"feed_emoji":"🔭","tokens_out":6663,"duration_ms":52973,"temperature":0.7,"pith_summary":"This paper asks whether the MIRI/LRS transmission spectrum of K2-18b actually shows any spectral features. Fitting six nested Gaussian models to the published spectrum, the author finds that a flat line is an acceptable fit in five of six cases; the only model that beats the flat line, two Gaussians fixed at 7 and 8.8 microns, does so weakly, with ln(B)=1.21, about a 2-sigma preference. This contradicts an earlier claim of a 3.4-sigma detection of spectral features and weakens the case that dimethyl sulfide or dimethyl disulfide has been detected in this dataset. A flat spectrum matters because biosignature claims require ruling out simpler, featureless explanations.","feed_headline":"No strong spectral features found in K2-18b's MIRI spectrum","feed_subtitle":"A flat line fits the JWST data; only weak 2-sigma preference for gas features remains.","key_machinery":"The central object is the Gaussian feature model $\\delta_\\lambda(A,\\mu,\\sigma_m,c)=A\\exp(-(\\lambda-\\mu)^2/(2\\sigma_m^2))+c$ combined with a constant flat-line model $\\delta_\\lambda(c)=c$. The argument runs on nested model comparison: because the flat line is a subset of every Gaussian model, the difference in log Bayesian evidence equals the log Bayes factor, judged on the Jeffreys scale. MultiNest nested sampling supplies the evidence values, and equivalent-$\\sigma$ conversion translates the Bayes factor into the familiar but potentially misleading '2-$\\sigma$' language.","core_discovery":"The central claim is that the K2-18b MIRI/LRS transmission spectrum does not contain statistically significant spectral features. The author reanalyzes the same binned spectrum presented in the earlier detection paper and compares six Gaussian-based models against a constant flat line using Bayesian evidence. Five of the six Gaussian models fall into the 'no evidence' category on the Jeffreys scale, and the most favorable model—positive and negative Gaussians fixed at 7 and 8.8 microns—has ln(B)=1.21, which is 'weak evidence' and roughly a 2-sigma preference. The reduced chi-squared values around 1 show the flat line fits the data well.","pith_inferences":["Editorial inference: the same flat-line-versus-Gaussian test could be applied to K2-18b's NIRISS/SOSS and NIRSpec/G395H spectra; if those also prefer a flat line, the case for any molecular detection in this planet's transmission spectra weakens further.","Editorial inference: an independent reduction of the raw MIRI/LRS time series would be the cleanest check, because the flat-line result depends on the error bars of the published binned spectrum.","Editorial inference: the method generalizes: any claimed exoplanet spectral feature can be required to beat a flat line in a nested Bayesian comparison before being interpreted as a molecular detection."],"forward_implications":["A flat-line description of the K2-18b MIRI/LRS spectrum is a valid baseline; the claimed 3.4-sigma feature detection is not reproduced by a model-agnostic search.","Atmospheric abundances of DMS/DMDS derived from this dataset carry much less weight than the original publication suggests.","Molecule-specific, fixed-wavelength models can inflate the statistical evidence for a detection relative to agnostic feature searches.","Future JWST observations of K2-18b in transmission should be designed with enough precision to distinguish a genuinely featureless spectrum from weak 2-sigma bumps."],"supporting_citations":[{"why":"Supplies the MIRI/LRS transmission spectrum and the 3.4-sigma detection claim that this paper re-tests.","marker":"N. Madhusudhan et al. 2025"},{"why":"Defines the Jeffreys evidence scale and the equivalent-sigma conversion used to interpret the Bayes factors.","marker":"R. Trotta 2008"},{"why":"Provides MultiNest nested-sampling evidence computation used for all model comparisons.","marker":"F. Feroz et al. 2009; J. Buchner et al. 2014"},{"why":"Establishes the Gaussian feature-search approach for transmission spectra of small planets that this paper follows.","marker":"E. M. May et al. 2023"},{"why":"Another precedent for searching spectra for Gaussian-shaped features independent of specific molecules.","marker":"JWST Transiting Exoplanet Community Early Release Science Team et al. 2023"},{"why":"Shows ethane should be more detectable than DMS at the proposed biogenic flux, motivating the doubt about the molecular claim.","marker":"S.-M. Tsai et al. 2024"}],"fun_headline_variants":["K2-18b's MIRI spectrum shows no strong spectral features","Flat line fits K2-18b's JWST data better than gas features","Reanalysis: no solid evidence for spectral features in K2-18b","K2-18b spectrum falls flat, gas feature detections weaken","MIRI spectrum of K2-18b: flat line wins over Gaussian features"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The conclusion rests on the accuracy of the published binned spectrum's error bars; if they are underestimated, real features could hide, and if overestimated, phantom flatness could appear.","fun_headline_variants_meta":{"raw":{"variants":["K2-18b's MIRI spectrum shows no strong spectral features","Flat line fits K2-18b's JWST data better than gas features","Reanalysis: no solid evidence for spectral features in K2-18b","K2-18b spectrum falls flat, gas feature detections weaken","MIRI spectrum of K2-18b: flat line wins over Gaussian features"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000245,"raw_usage":{"total_tokens":1506,"prompt_tokens":885,"completion_tokens":621,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":501,"completion_tokens_details":{"reasoning_tokens":518}},"tokens_in":501,"tokens_out":621,"duration_ms":5424,"temperature":1.0,"reasoning_tokens":518,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T11:14:38.932436+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A re-reduction of the raw MIRI/LRS time series that yields materially different transit depths near 7 and 8.8 microns, or an injection-recovery test showing that the method fails to recover a simulated feature at the claimed amplitude, would overturn or confirm the flat-line conclusion.","supporting_citations":[{"cited_title":"New Constraints on DMS and DMDS in the Atmosphere of K2-18 b from JWST MIRI","cited_arxiv_id":"2504.12267","evidence_quote":"Supplies the MIRI/LRS transmission spectrum and the 3.4-sigma detection claim that this paper re-tests."}],"review_version":1}