{"id":"76ae8f1e-7247-4fb7-8439-03c0cd276332","arxiv_id":"2607.22512","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Routine CEM I cement characterization supports useful cross-producer inference of 28-day strength, strength class, and water demand, but does not resolve early-strength N/R designation or producer-specific process effects.","lead":"This study tests how well routine quality-control measurements of ordinary Portland cement — chemistry, fineness, particle-size data — can predict strength class, water demand, and 28-day compressive strength across 23 European producers. It finds useful cross-producer prediction for strength and water demand, but shows that the early-strength N/R label and producer-specific strength variation are not fully recoverable from routine data alone.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Potential selection bias in the PSD-complete subset (n=181) used for all headline 28-day strength and producer-holdout results; no missingness comparison is provided.","rationale":"Our review focused on the empirical core: cross-producer inference of 28-day strength. The headline MAE and the producer-holdout results that substantiate transferability are computed on the PSD-complete subset only. The 36% missingness is the largest feature-level gap in the CEM I data (Table 2) and the paper gives no evidence that the observed subset is representative. Since PSD is the second descriptor family in the best-performing 'Oxides+PSD' configuration, and since the producer-holdout analysis necessarily includes only producers with enough PSD-complete samples (11 of 20), non-random missingness could both bias the performance estimates and change the set of producers on which transfer is evaluated. This is more directly damaging to the central claim than the stationarity concern, because even if producer effects were perfectly stationary, the subset selection alone could invalidate the generalization. The paper's own native missing-value handling for tree models makes the full-data regression a trivial addition, so the omission is a real gap rather than a computational constraint. We agree with the reader's CONDITIONAL verdict: the paper is honest and well-executed, but should be required to test the missingness assumption before the numeric claims are taken as population-level. Our concern is narrower than the reader's weakest assumption (which included time stationarity), hence partial agreement.","tokens_in":28436,"tokens_out":8974,"duration_ms":85160,"concrete_test":"Using the deposited Zenodo dataset, (1) compare the 181 PSD-complete CEM I records to the 130 PSD-missing records on all reported features (oxides, Blaine, density, 28-day strength, water demand, producer, and any collection date fields) via standardized mean differences and a multivariate balance test; (2) re-run the 5-fold CV XGB regression of 28-day strength on all 311 CEM I records with native missing-value handling (same hyperparameters as Table 6) and compare MAE/R2 to the n=181 result. If any |standardized mean difference| > 0.5 or the full-data XGB MAE deviates from 2.71 MPa by more than 5%, the headline regression and transfer numbers should be re-reported on the full cohort with missingness indicators or multiple imputation.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central quantitative claims for 28-day strength rest exclusively on the PSD-complete CEM I cohort of n=181 (Tables 6, 7, 9; Sections 3.5.2 and 3.6.2). This cohort is obtained by dropping 130 of 311 CEM I records with jointly missing PSD parameters (36%, Table 2). The paper does not compare the complete and incomplete records on any covariate or target. The stated reason for missingness — 'PSD measurement was not routinely conducted for all samples and, when performed, the parameters were recorded separately' (Section 2.2.3) — indicates an administrative, not random, mechanism. If PSD availability correlates with producer, time period, or cement characteristics, then the Oxides+PSD performance (XGB MAE 2.71 MPa, R2=0.638) and especially the producer-holdout results (Table 9: pooled MAE 2.94–3.20 MPa across 11 producers and 164 samples) are conditional on a selected subsample and may not generalize to the full CEM I population. The paper's own tree models natively handle missing PSD and are used for strength-class and water-demand targets on all 311 records (Table 5), but no analogous full-data regression for 28-day strength is reported. This is a direct, addressable threat to the central cross-producer inference claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper analyzes a 476-sample, 23-producer, one-laboratory cement database spanning 27 years, asking how much routine CEM I characterization (oxides, Blaine, PSD, derived Bogue phases, equivalent alkali) can infer about 28-day compressive strength, EN 197-1 strength class, water demand, and the N/R early-strength designation. Using OLS/ENET, random forest, XGBoost, and LightGBM with 5-fold CV and producer-holdout experiments, it reports CEM I 28-day strength MAE of 2.71–2.97 MPa with R² up to 0.638 on the PSD-complete subset (n=181), strength-class balanced accuracy 0.894, water demand R²=0.713, and producer-holdout pooled MAE 2.94–3.20 MPa for 11 producers at n≥5. The N/R label is shown to be recoverable only as a population-level tendency, not as a separable class. The paper is unusually explicit about several limitations, but the central cross-producer quantitative claims rest on a PSD-complete subsample whose representativeness is not established.","tokens_in":28674,"tokens_out":5927,"duration_ms":61690,"significance":"If the results hold, the paper provides a useful, quantitative benchmark for the information content of routine cement QC data and a careful decomposition of what transfers across producers versus what remains producer-specific. The study has clear strengths: a relatively large multi-producer dataset with standardized testing, explicit disclosure that randomized folds measure within-population prediction rather than transfer, explicit acknowledgment that Na2Oeq mainly tracks K2O, a physically motivated treatment of N/R as an inference limit, and deposited data and code. However, because the headline 28-day strength and producer-holdout results are all conditional on the 181-sample PSD-complete cohort, the main cross-producer generalization claim is not yet fully supported.","major_comments":[{"comment":"The central 28-day strength and producer-holdout results are obtained exclusively on the PSD-complete CEM I subset (n=181), after dropping 130 of 311 records (36%) with jointly missing PSD parameters. The paper states that PSD was not routinely measured and was recorded separately, which indicates a non-random, administrative missingness mechanism, but it never compares the complete and incomplete records on covariates, targets, producers, or sampling period. If PSD availability correlates with producer, time, or material characteristics, the reported MAE 2.71 MPa and producer-holdout pooled MAE 2.94–3.20 MPa may not generalize to the full CEM I population. Please add a missingness analysis and report a 28-day strength regression on all 311 CEM I records using the native missing-value handling already used in Table 5, or an appropriate imputation, including producer-holdout performance i","section":"§2.2.3, Table 2; §3.5.2, Tables 6, 7, 9"},{"comment":"The data span 27 years, and Section 2.1 acknowledges variability due to plant changes and measurement-equipment revisions, but no epoch or campaign information is used in the producer-holdout tests. If producer labels are confounded with time periods or measurement generations, the 'cross-producer' transfer results in Table 9 may partly reflect cross-era transfer rather than transfer across independent producers. Please assess whether sampling years are balanced across producers and PSD availability, and consider adding a temporal covariate or temporally blocked cross-validation as a robustness check. At minimum, state explicitly whether any temporal covariate exists in the dataset.","section":"§2.1, §3.6.2"}],"minor_comments":[{"comment":"The 'Full' cohort is listed as n=468, whereas Table 1 reports 476 total samples and 311 CEM I. The difference is not explained; please clarify whether eight records are excluded from the strength-class analyses and why.","section":"Table 5"},{"comment":"The text refers to a '200-sample complete subset' for PSD, while Table 6 uses n=181 after target availability, complete-case filtering, and the 5σ screen. The sequential reduction from 200 to 185 to 181 is described later in Section 3.5.1, but it would be clearer to state this reduction at the point of the first mention.","section":"§2.2.3, Table 6"},{"comment":"The Bogue correction assumes a fixed 3 wt% gypsum addition; the sensitivity of the derived Bogue phases to this assumption is not explored. A short sensitivity statement would help, especially because Table A.1 reports per-sample changes from this correction.","section":"Appendix A"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is careful and competent, and the main obstacle is the unexamined PSD-complete subsample. The issue is directly addressable by adding a missingness comparison and full-data regression; if those analyses do not change the conclusions, I would be willing to accept the paper after revision. I do not see grounds for rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's the short version: this is a competent, honest ML study on a real multi-producer cement dataset, and its main qualitative conclusions survive a skeptical read. What's actually new is the transfer test: holding out entire producers and showing pooled absolute errors stay in the same 2.9–3.2 MPa range while R2 drops and varies by producer. That distinction is made cleanly, and the paper is careful to separate within-population CV from producer transfer. It also ships the data and scripts, which makes the claims checkable.\n\nThe other results—fineness dominates, oxides add comparable signal, Na2Oeq is mostly K2O, N/R is not a separable class—are not surprising given the domain literature, but they are now backed by a 27-year, 23-producer, single-lab database instead of single-plant case studies. The Bogue calculation is transparently documented, including the fixed 3 wt% gypsum assumption that is a real simplification.\n\nSoft spots, in order of severity. First, the headline 28-day numbers all come from the PSD-complete cohort (n=181), which drops 130 of 311 CEM I records. The paper gives no comparison of the complete and incomplete records on any covariate or target. Missingness is described as administrative, but 'PSD not routinely done' can easily correlate with producer or time period, and the producer-holdout results depend on which producers happen to have PSD. This is addressable—run a missingness table, or re-run the 28-day regression on all 311 with tree models that already handle missing PSD—but right now it is the main threat to the exact numbers, not the qualitative claims.\n\nSecond, the producer holdout at n≥5 includes two producers with five samples each; the paper admits individual R2 is too unstable there, and pooled R2 drops from 0.52 to 0.36 as the threshold rises. I don't think this invalidates the MAE transfer conclusion, but it should be reported with that sensitivity made more prominent. Third, the N/R non-separability conclusion would be stronger with a CV protocol for the logistic classifier; the 75% balanced accuracy is presented without error bars. Minor: the Bogue/gypsum assumption is known to disagree with XRD on the C3S/C2S split, and the paper shows that in Table A.2 but still uses Bogue as a primary feature set—that's okay for prediction but should not be read as phase quantification.\n\nWho is this for: anyone working on virtual qualification or cross-plant ML in cement, and the dataset deposit alone is valuable. I'd send it to peer review, with a request for the missingness comparison and a full-data regression as a revision condition, not a desk reject. If the missingness check comes back clean, the paper's core message holds.","headline":"A careful, honest ML study on a multi-producer cement dataset; the main qualitative claims survive, but the PSD-missing subsample needs a missingness check before the headline numbers are fully trusted.","tokens_in":29324,"tokens_out":2150,"would_cite":true,"duration_ms":21008,"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":"Routine cement quality-control data carry enough transferable information to predict 28-day strength, strength class, and water demand across independent producers; the N/R early-strength label is only a population-level tendency.","keywords":["Portland cement","machine learning","compressive strength","model transferability","chemical composition","particle size distribution","Bogue phases","cement fineness"],"falsifier":"Measure a held-out producer whose clinker uses atypical sulfate speciation or grinding practice (for example, high gypsum substitution or a different aluminate polymorph distribution) and check whether 28-day strength predictions fall outside the 2.94–3.20 MPa MAE band observed here; a systematic excursion beyond roughly 4 MPa would indicate the cross-producer transfer claim is limited to the sampled process envelope. Alternatively, add XRD-based phase and sulfate-form variables to the N/R classification: if balanced accuracy rises well above the reported ~0.75, the conclusion that routine dat","tokens_in":28184,"feed_emoji":"🏗️","tokens_out":6248,"duration_ms":61181,"temperature":0.7,"pith_summary":"The paper argues that routine quality-control measurements on ordinary Portland cement—oxide chemistry, Blaine fineness, and particle-size descriptors—carry enough information to predict a cement's 28-day compressive strength, strength class, and water demand even for producers not seen during training. On a dataset of 476 samples from 23 European producers collected in one laboratory over 27 years, prediction errors for 28-day strength stay around 2.7–3.2 MPa, and strength-class balanced accuracy reaches about 0.89. The one clear boundary is the N/R early-strength designation: the data support only a population-level tendency, not a physically separable class, because early strength can be produced through multiple combinations of fineness, sulfate–alkali chemistry, phase assemblage, and plant practice. If these results hold, routine cement characterization could support preliminary performance screening and cross-producer comparison without waiting for full 28-day tests.","feed_headline":"Cement QC data transfers across plants: 28-day strength within ~3 MPa","feed_subtitle":"Routine chemistry and fineness data transfer across 23 producers, with one clear limit: early-strength labels.","key_machinery":"The load-bearing machinery is the multi-producer dataset itself: 476 cement records from 23 European producers, all tested in one laboratory over 27 years under DIN EN methods, which lets within-producer covariation be separated from cross-producer signal. On top of this sit three tools: a preprocessing pipeline that computes corrected Bogue phase fractions (calculated estimates of clinker minerals derived from oxide chemistry), equivalent alkali, and clinker moduli; explainable gradient-boosted tree models with Shapley attribution to quantify group-level feature reliance; and producer-holdout validation, in which an entire producer is excluded from training to test transferability. The key","core_discovery":"The central finding is that the information routinely collected for cement quality control is sufficient for practically useful inference of CEM I performance across independent producers, with a specific boundary. In 5-fold cross-validation, the best model predicts 28-day compressive strength with mean absolute error 2.71 MPa (R² = 0.638) using oxides plus particle-size descriptors; producer-holdout tests across 11 unseen producers with at least five samples give pooled MAE 2.94–3.20 MPa and R² of 0.48–0.52. Strength class is recovered with balanced accuracy 0.894 and water demand with R² = 0.713. Fineness—measured either as Blaine or as a compact set of Rosin–Rammler percentiles—is the str","pith_inferences":["If the same methodology were applied to blended cements (CEM II/III), the clinker-fraction signal would likely dominate and cross-producer transfer errors could differ markedly; the paper does not claim its results extend to those types.","A direct test of the K2O signal would be to manipulate the sulfate-to-alkali balance within a single plant: if strength recovers when potassium sulfate availability is adjusted, the association is mechanistic; if not, it is a proxy for other producer-level differences.","The interchangeability of Blaine and compact PSD suggests that historical databases lacking PSD measurements can still support retrospective strength models as long as Blaine and oxides are present—an implication the paper leaves implicit.","The pooled producer-holdout errors invite a practical extension: an online calibration scheme in which a new producer contributes a small number of samples to refine a population model, testing whether the transfer gap closes with modest local data."],"forward_implications":["A producer or purchaser could use routine quality-control data alone to screen 28-day compressive strength within roughly 3 MPa before the full 28-day test is complete, and to flag cements whose strength class is likely mislabeled.","Because strength class is recovered at about 0.89 balanced accuracy and water demand at R² ≈ 0.71, routine characterization could support automated data-quality checks on cement certificates and faster inter-laboratory comparison.","Blaine fineness and a compact particle-size distribution representation are largely interchangeable in this dataset, so legacy databases with only Blaine can support the same strength-inference models without full PSD measurements.","The consistent negative K2O/Na2Oeq association with 28-day strength, treated as a population-level monitoring signal, could feed clinker-design feedback aimed at limiting alkali-related strength loss.","The N/R early-strength designation cannot be reliably inferred from routine data; any system that tries to predict or certify it must add speciation measurements such as sulfate form or aluminate polymorphs, or accept a non-deterministic, population-level label."],"fun_headline_variants":["Routine cement QC data predicts 28-day strength within ~3 MPa across producers","Cement QC info transfers: 28-day strength MAE ~3 MPa, early-strength weak","Cement characterization: fineness and oxides recover strength, water demand","Cement QC: 2.71 MPa MAE for 28-day strength, but early-strength is fuzzy","Routine cement data: strength and water demand recoverable across 23 producers"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The dataset of 23 producers collected over 27 years is treated as a representative population in which producer effects and measurement conditions are stationary; if producer labels are confounded with time periods, equipment changes, or unrecorded process revisions, the cross-producer transfer results could reflect hidden batch effects rather than transferable chemistry–strength relationships.","fun_headline_variants_meta":{"raw":{"variants":["Routine cement QC data predicts 28-day strength within ~3 MPa across producers","Cement QC info transfers: 28-day strength MAE ~3 MPa, early-strength weak","Cement characterization: fineness and oxides recover strength, water demand","Cement QC: 2.71 MPa MAE for 28-day strength, but early-strength is fuzzy","Routine cement data: strength and water demand recoverable across 23 producers"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00104,"raw_usage":{"total_tokens":4259,"prompt_tokens":842,"completion_tokens":3417,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":586,"completion_tokens_details":{"reasoning_tokens":3304}},"tokens_in":586,"tokens_out":3417,"duration_ms":24205,"temperature":1.0,"reasoning_tokens":3304,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T04:29:24.064185+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure a held-out producer whose clinker uses atypical sulfate speciation or grinding practice (for example, high gypsum substitution or a different aluminate polymorph distribution) and check whether 28-day strength predictions fall outside the 2.94–3.20 MPa MAE band observed here; a systematic excursion beyond roughly 4 MPa would indicate the cross-producer transfer claim is limited to the sampled process envelope. Alternatively, add XRD-based phase and sulfate-form variables to the N/R classification: if balanced accuracy rises well above the reported ~0.75, the conclusion that routine dat","supporting_citations":[],"review_version":1}