{"id":"355b452c-530e-450e-90b1-921478114043","arxiv_id":"2511.19277","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Climate TRACE provides monthly, asset-level greenhouse gas emissions estimates for essentially all anthropogenic sectors worldwide — the first global dataset with this level of detail.","lead":"Climate TRACE is an open-access dataset that estimates greenhouse-gas emissions for individual facilities — power plants, factories, cattle operations — worldwide, with monthly updates. It claims to be the first globally comprehensive source-level emissions inventory, designed to support emissions monitoring, corporate accounting, and mitigation planning.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'asset-level for all sectors' claim is contradicted by the ~26% of emissions distributed via proxies, not assigned to actual sources.","rationale":"I identify a discrete, load-bearing inconsistency in the paper's central claim: the dataset is advertised as providing asset-level emissions estimates for all emitting sectors, yet the methodology explicitly allocates a substantial share of emissions (at least 26% globally, per the 74% facility-level coverage figure) via proxy-based spatial disaggregation rather than source-level attribution. This directly undermines the 'first global asset-level dataset' claim, because a significant portion of the data are not asset-level at all. The reader's weakest assumption focused on the global validity of models trained in high-income countries and the remainder allocation as one combined risk. I agree with the latter part but consider it the primary issue, since even if models generalized perfectly, the remainder problem would still mean the dataset fails to provide source-level estimates for all sectors. The paper already discloses the 74% figure, but it does not reconcile this with the 'all sectors' claim, and it does not provide subsector-level remainder fractions in the main text. A conditional acceptance is appropriate: the headline claims must be revised to specify the fraction of emissions that are truly source-level, and the remainder fractions by sector must be made public. This is a concrete, testable condition, not a subjective objection.","tokens_in":22947,"tokens_out":7086,"duration_ms":75958,"concrete_test":"Using the methodology documents (e.g., Table 2 of the 'Spatial Disaggregation of Remainder Emissions' document), compute for each subsector the fraction of emissions that are 'remainder' (i.e., not directly assigned to identified assets) in the latest data release. Also compute the global fraction of remainder emissions and reconcile it with the main text's 74% facility-level coverage figure. If any sector's remainder fraction exceeds 10%, or if the global remainder exceeds 25%, the claim that the dataset provides asset-level estimates for all sectors is untenable without substantial qualification. Publish these fractions by sector alongside the dataset to allow independent verification.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The paper's central claim—that Climate TRACE is the first global dataset to provide asset-level emissions estimates for all anthropogenic sectors—is not supported by its own methodology. Supplementary Note 1.2 describes a spatial disaggregation protocol in which, when a chosen country-level estimate exceeds the sum of Climate TRACE asset-level estimates, the remainder is not attributed to any individual source but is spread across space using proxies such as population or nightlights. The main text (Section 3.3) states that monthly facility-level estimates cover only 74% of global anthropogenic emissions; the remaining ~26% are, by construction, not source-level. Several subsectors reported in Supplementary Note 1.3 are dominated by such remainders (e.g., 'other manufacturing' at 2.89% of global emissions, 'other energy use' at 2.71%, 'other agricultural soil emissions' at 2.34%). For these subsectors, the 'asset-level' estimates are effectively country totals downscaled to administrative boundaries or grid cells, not data tied to individual facilities. Consequently, the dataset is not truly a global source-level inventory for all sectors, and the promised ability to identify high-emitting facilities and under-reporters is only meaningful for the subset of emissions directly linked to identified assets. This is a structural mismatch between the headline claim and the actual data architecture, not merely a validation gap.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript presents Climate TRACE, an open-access global greenhouse gas emissions dataset and platform. It first offers a structured comparison of existing global emissions datasets across accuracy, coverage/resolution, update frequency, and usability, identifying gaps that motivate the new dataset. It then describes the Climate TRACE framework: synthesizing existing datasets (e.g., EDGAR, CEDS, UNFCCC) and generating new sector-specific estimates, often with machine learning from satellite imagery, to produce emissions estimates for individual facilities (e.g., power plants) and monthly, globally resolved emissions from 2021 onward. The paper claims this is the first global dataset providing asset-level estimates for all anthropogenic emitting sectors, with a two-month reporting lag and monthly updates, and it highlights applications in company-level under-reporting detection, subnational inventories, and supply-chain accounting. The paper also reports global emissions trend analyses from the dataset. Limitations are acknowledged regarding ground-truth scarcity and variable confidence across sectors.","tokens_in":23347,"tokens_out":2667,"duration_ms":31237,"significance":"If the claims as stated were fully supported, this would be a substantial contribution: a single, open, monthly, globally comprehensive emissions inventory spanning multiple gases and sectors, with facility-level detail for a large share of emissions, would be of clear value to governments, companies, and researchers. The paper ships an actual dataset and platform, with detailed sector methodology documents, reproducible R scripts for its trend analyses, and external validation comparisons for components such as power plant emissions. These are real strengths. However, the central novelty claim—'asset-level estimates for all emitting sectors'—is materially overstated by the paper's own methodology, and the accuracy evidence is uneven. The contribution is still defensible if the claims are carefully qualified to distinguish directly asset-linked emissions from spatially proxied remainder emissions, and if uncertainty data are made openly available rather than 'upon request'.","major_comments":[{"comment":"The central claim that Climate TRACE provides 'asset-level emissions estimates for all emitting sectors' (Abstract; also Discussion, Section 4) is contradicted by the paper's own spatial disaggregation protocol. Supplementary Note 1.2 states that when a country-level estimate exceeds the sum of asset-level estimates, the remainder is distributed using proxies such as population and nightlights, not assigned to identified sources. Section 3.3 reports that monthly facility-level estimates cover only 74% of global anthropogenic emissions, implying ~26% is not source-level by construction. Several subsectors explicitly rely on such remainders (Supplementary Note 1.3: 'other manufacturing' 2.89%, 'other energy use' 2.71%, 'other agricultural soil emissions' 2.34%). The manuscript must either revise the headline claim to 'asset-level for a large subset of emissions, with the remainder downscal","section":"Abstract; Section 3.3; Supplementary Note 1.2"},{"comment":"Transparency and reproducibility are central motivations, but quantitative uncertainty metrics are 'available upon request' rather than directly downloadable. This weakens the open-data claim and prevents independent users from applying the confidence information that the paper says is a key improvement over other datasets. The manuscript should make uncertainty/confidence data part of the publicly downloadable product, or explain why this is not possible. The 'available upon request' language is also inconsistent with the Data and Materials Availability statement that all data are available for download.","section":"Section 3.1; Supplementary Note 4.1"},{"comment":"Validation of power plant emissions, which the paper uses as a flagship example, does not demonstrate global generalizability. The ML models are trained on reported generation data from the USA, Europe, and Australia, and the main validation compares against Vulcan, a US inventory built on EPA reported data—the same reporting system used in training. The paper cites an external study (ref. 76) but does not provide out-of-sample validation for the plants outside these regions where the bulk of global emissions occur. The claim that models 'are trained on in situ ... data from the United States, Europe, and Australia' and then applied worldwide requires explicit evidence of transfer performance in low- and middle-income countries, or a clear statement that the accuracy is unverified there. This is load-bearing because the paper's actionability claim relies on identifying high-emitting faci","section":"Supplementary Notes 2.1.1 and 2.1.2"},{"comment":"The manuscript uses EDGAR and CEDS both as inputs (e.g., for remainder allocation and co-pollutant ratios) and as comparison references for validation. This creates a circularity concern for the comparative validation: agreement with EDGAR/CEDS may partly reflect shared input data rather than independent confirmation. The paper should explicitly identify which validation comparisons are independent of the input datasets and, where circularity exists, state that the comparison only checks consistency, not accuracy. This is not a fatal flaw, but it affects how strongly the accuracy claims can be framed.","section":"Supplementary Note 1.2; Supplementary Note 3.0"}],"minor_comments":[{"comment":"The Abstract says 'most anthropogenic emitting sectors' but the Discussion (Section 4) says 'all emitting sectors' and 'for all anthropogenic emitting sectors.' These are inconsistent; the qualified version is more accurate.","section":"Abstract vs. Discussion"},{"comment":"The third contribution ('regularly updated and openly accessible dataset') would be strengthened by explicitly stating the spatial coverage caveat (asset-level for ~74% of emissions; remainder spatially downscaled).","section":"Section 1.2 / Contributions"},{"comment":"Equations 1 and 2 are presented as the framework for 'all Climate TRACE subsectors,' but Supplementary Note 1.3 describes an implicit estimation approach that subtracts one dataset from another (e.g., EDGAR metals minus iron/steel/aluminum) and then spatially disaggregates. This does not obviously follow Equations 1 and 2; please clarify how the implicit method maps onto the activity/emission-factor framework.","section":"Section 5.2.2"},{"comment":"The cattle regression validation reports Spearman r from 0.32 to 0.8, with weaker correlations for eastern US beef. This is a useful honest report, but the manuscript should note the implication for emissions estimates: a 0.32 correlation indicates large uncertainty in a substantial share of cattle emissions, not just in one regional model.","section":"Supplementary Note 2.2.2"},{"comment":"The axis labels and units are not visible in the figure as provided; please ensure the final version includes clearly labeled axes and a legend.","section":"Supplementary Figure S1"}],"recommendation":"major_revision","confidential_remarks":"The manuscript describes a genuinely useful and publicly available dataset, but the flagship claim of 'first global asset-level inventory for all sectors' is overstated by the paper's own methodology (~26% is proxy-allocated), and validation is concentrated in high-income regions. These are fixable with clarified claims, a sector-level asset/remainder breakdown, and public uncertainty data. I would not reject, but the revision needs to be substantive rather than cosmetic."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. First, Climate TRACE is a genuinely useful artifact: open, monthly, global, covering all IPCC sectors, with subnational aggregations and a real platform. Second, the paper's central claim—\"first global dataset to provide asset-level emissions estimates for all emitting sectors\"—is not supported by its own methodology. Section 3.3 says monthly facility-level data cover 74% of global emissions; Supplementary Note 1.2 explains the rest is remainder emissions allocated with proxies like population and nightlights. For subsectors like \"other manufacturing\" (2.89%) and \"other energy use\" (2.71%), the so-called asset-level estimates are country totals downscaled to grid cells.\n\nWhat's good: the dataset is more than vaporware. The survey of the data landscape (Table 2) is useful. The synthesis is genuinely new at this scale—no prior dataset with this sectoral breadth, monthly cadence, and 74% source-level coverage. The power plant ML work is a real attempt, and the cattle database with satellite detection moves beyond simple inventory copying. Publishing methodology documents on GitHub is a plus.\n\nSoft spots are real but not disqualifying. (1) Claim-reality gap: \"asset-level for all sectors\" should be tempered to \"asset-level where assets can be identified, with country-level totals elsewhere.\" (2) Validation circularity: EDGAR/CEDS appear as both inputs and comparison references; power plant models trained on US/EU/Australia data are applied worldwide, with limited independent validation. The paper is honest that accuracy cannot be directly assessed, but the comparative validations sometimes compare against derivatives of the same inputs. (3) Uncertainty data are \"available upon request\"—that is not transparency in a data paper. These should be downloadable with the dataset.\n\nThis deserves a serious referee. It's an important infrastructure paper, not a polished accuracy claim. A good review could push the authors to fix the framing and release the uncertainty metrics. I'd bring it to reading group as a case study in how dataset papers should—and shouldn't—frame coverage claims.","headline":"Valuable open dataset, but the 'asset-level for all sectors' headline doesn't survive contact with the paper's own methods: a quarter of global emissions are proxy-downscaled remainders, not facility-level data.","tokens_in":23929,"tokens_out":2169,"would_cite":true,"duration_ms":23081,"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":"The paper claims Climate TRACE is the first global dataset to provide source-level emissions estimates for all major anthropogenic sectors, updated monthly.","keywords":["greenhouse gas emissions","emissions monitoring","asset-level inventory","Climate TRACE","satellite remote sensing","machine learning","carbon accounting","spatial disaggregation"],"falsifier":"Take a random global sample of facilities outside the training regions, such as power plants in Africa, South America, and Southeast Asia with independently audited hourly generation and fuel records, and compare Climate TRACE monthly estimates to those records. If errors outside training regions are systematically larger and more biased than errors in the US, Europe, and Australia, the global-validity claim fails.","tokens_in":22890,"feed_emoji":"🛰️","tokens_out":5122,"duration_ms":48899,"temperature":0.7,"pith_summary":"Existing global greenhouse gas datasets are each missing at least one feature needed for actionable climate monitoring: independent validation, global coverage, fine spatial or temporal resolution, frequent updates, or ease of use. The paper argues that these gaps can be closed by a single framework that combines the best existing data and fills the rest with sector-specific estimates. Its central claim is that the resulting dataset, Climate TRACE, is the first to provide globally comprehensive emissions estimates for individual sources across essentially all anthropogenic sectors, with monthly updates and a two-month lag. If true, this makes facility-level emissions visible and comparable worldwide, enabling detection of high emitters and under-reporters and giving subnational governments a ready-made baseline for mitigation.","feed_headline":"Emissions monitoring reaches every sector, down to single facilities","feed_subtitle":"Monthly, open-access source-level estimates could reveal under-reporters and power local mitigation.","key_machinery":"The load-bearing identity is the capacity-activity-emission-factor chain, which lets the framework estimate emissions from things that can be observed remotely. The novel machinery is the synthesis pipeline: candidate datasets are screened for reliability, combined or adapted via data-informed disaggregation and temporal imputation, and supplemented with newly generated estimates from machine-learning models trained on satellite imagery and in situ data. A spatial-disaggregation rule treats country-level totals as a lower bound and distributes unlocated emissions according to sector-specific proxies, ensuring that every administrative unit and grid cell receives a complete estimate.","core_discovery":"Climate TRACE is built around a universal accounting identity: for each source, activity equals capacity times capacity factor, and emissions equal activity times emission factor. Sector-specific methods estimate these components using machine learning on satellite imagery for power plants and cattle operations, statistical models for oil and gas, and flight-level fuel calculations for aviation, while remaining gaps are filled by disaggregating country-level totals using proxies such as population and nightlights. The assembled dataset covers all major anthropogenic sectors, provides monthly asset-level estimates for roughly 74% of global emissions, and offers subnational aggregations for ne","pith_inferences":["If the trained models' high-income training data do not generalize, cross-country source-level rankings could be biased; a natural test is to compare estimates against independently audited facilities in lower-income regions.","The lower-bound assumption on country totals means that in countries where asset-level detection is incomplete, remainder emissions are assigned by proxies, so local emission hotspots could be misplaced even if national totals are right.","Monthly, source-level data could enable near-real-time attribution of emission changes to specific shutdowns, strikes, or policy interventions, something annual inventories cannot provide.","A direct extension would be to use the same framework to produce forward-looking emissions projections tied to capacity changes, which the paper does not attempt."],"forward_implications":["Individual power plants, factories, cattle operations, airports, and other facilities worldwide become individually monitorable, allowing comparisons that were previously possible only in a few countries.","Self-reported emissions can be checked against independent estimates, potentially exposing systematic under-reporting at the company or sector level.","Subnational governments, including urban areas and low-capacity countries, gain emissions baselines and can set and track mitigation targets without building their own inventories.","Atmospheric inversion models can use source-level priors, which the paper shows improves alignment between bottom-up and top-down methane estimates for cattle.","Monthly updates with a two-month lag make it possible to evaluate mitigation policies on the timescale of a year rather than a multi-year inventory cycle."],"fun_headline_variants":["Open data reveals emissions for every major sector, down to single plants","Source-level emissions for 74% of global output, updated monthly","Every major sector, every major source: monthly emissions now open","Pinpoint emissions to individual facilities across all major sectors","Monthly source-level emissions data for every major sector, open access"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that models trained on data from a few high-income countries, combined with proxy-based allocation of unlocated emissions, remain accurate enough when applied to every country and facility worldwide.","fun_headline_variants_meta":{"raw":{"variants":["Open data reveals emissions for every major sector, down to single plants","Source-level emissions for 74% of global output, updated monthly","Every major sector, every major source: monthly emissions now open","Pinpoint emissions to individual facilities across all major sectors","Monthly source-level emissions data for every major sector, open access"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001174,"raw_usage":{"total_tokens":4671,"prompt_tokens":704,"completion_tokens":3967,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":448,"completion_tokens_details":{"reasoning_tokens":3881}},"tokens_in":448,"tokens_out":3967,"duration_ms":27214,"temperature":1.0,"reasoning_tokens":3881,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T20:30:26.116770+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a random global sample of facilities outside the training regions, such as power plants in Africa, South America, and Southeast Asia with independently audited hourly generation and fuel records, and compare Climate TRACE monthly estimates to those records. If errors outside training regions are systematically larger and more biased than errors in the US, Europe, and Australia, the global-validity claim fails.","supporting_citations":[],"review_version":1}