Corpus 1,286 assessed · 1,187 scored · 648 reproduced ≥75 · 174 flagged ·∅ 73.9/100
← New search

Multiomic machine learning on lactylation for molecular typing and prognosis of lung adenocarcinoma.

Sci Rep · 2025
L1 83/100 PQI 94
Why this verdict

The main results reproduced: recomputed values matched the published ones within tolerance.

Reproduced on the brainbox compute brainarbeit.com
Scoring basis — itemised

Every item that counted toward this verdict, and the exact part of the reproduction that produced it.

Supporting (toward a concern)
Content-critical question only partially held
+2 pts
From: Q8 · Severity of the miss (overall human judgment) 🟡
Minor / cosmetic deviation
+1 pts
From: Q3 · Location of the main deviation 🟡
Minor / cosmetic deviation
+1 pts
From: Q4 · Cause of the deviation 🟡
Minor / cosmetic deviation
+1 pts
From: Q6 · Severity of the deviation 🟡
Concordant (toward reproduced)
Code + data deposited & functional
-2 pts
From: Data & code availability Available & functional
Total score +3
✓ What held up
  • Reported values were directly comparable
  • Reported values are derivable from the shared data
  • The central claim held under reproduction
What did not (or only partly)
  • 🟡Could not use the authors’ exact input data
  • 🟡A deviation arose in the data or preprocessing
  • 🟡A deviation was attributed to the published material
  • 🟡The deviation was non-trivial in magnitude
  • 🟡Overall, the reproduction showed a material discrepancy
How its reproducibility compares
83/100
Reproducibility score
0.5 SD above mean
vs. all fields · 1187 studies
🎯 Scores higher than 61% of all assessed papers rank 431 of 1187 scored

A 0–100 reproducibility-quality score from the per-question grades, shown as a z-score: standard deviations above (+) or below (−) the mean of comparable assessments.

Reproduction agent’s raw note

Described well enough -> 1:1 (within tolerance) on the headline prognostic claim. Reproduced the 9-gene lactylation Random Survival Forest risk model (authors' own code, Script/08_Machine_learning.R, seed=123/ntree=1000/nodesize=5/logrank) trained on TCGA-LUAD and validated on public GEO cohorts. All six reported time-dependent OS AUCs reproduce within +/-0.065 (GSE31210 0.810/0.754/0.756 vs 0.833/0.793/0.783; GSE13213 0.794/0.725/0.720 vs 0.730/0.683/0.668); all 9 hub genes present; RSF gives valid out-of-sample discrimination (C-index 0.736/0.699) and significant risk-group KM separation (GSE31210 p=2.6e-4). Graded partial/within-tol (not exact) for two honest reasons: (1) scope -- only the prognostic-model + validation half was attempted; (2) the TCGA training set is an approximation (the multiomic-complete TCGA subset + its normalization are NOT shipped; README says data 'available upon request'), so the RSF is not byte-identical even at seed=123. NOT attempted (out of scope, needs non-shipped intermediates): the MOVICS 10-algorithm multiomic clustering into 2 cancer subtypes CS1/CS2, and the ~12 downstream analyses (GSVA/ssGSEA, IOBR immune deconvolution, stemness/TMB/MSI, microenvironment, chemokine, oncoPredict drug-resistance, maftools mutation panels). No fabrication flag: validation data is public and the reproduced AUCs (some higher, some lower than reported) corroborate the claim. Data: TCGA-LUAD (UCSC-Xena GDC STAR FPKM-UQ + survival), GEO GSE31210 (GPL570, n=226), GSE13213 (GPL6480, n=117). Repo commit 2bbb5b53c005395e9da45d4f1f54b53283161a8d.

💻 Code ↗ 🗄 Data: GSE31210

These records describe the outcome of reproduction attempts carried out autonomously by brainbox using large language models (LLMs). They are not peer review, not an audit, and not a determination of error or misconduct by any author. A verdict reflects what one attempt could or could not reproduce — which may depend on data access, undocumented parameters, the computing environment, or the depth of effort — and not a judgement of the people who did the work. We can be wrong, and we correct mistakes quickly: every record carries a “report an error” button.

Assessment versions

Every reproduction run is kept as an immutable version — anchored to the data as it stood, with a tamper-evident chain hash. A rerun (e.g. after an author updates a deposit) adds a new version; the previous one stays on record.

  1. v1 current initial assessment Score 83
    assessed: 2026-06-14 ⛓ 2cc4883de7d5
✎ I am an author of this paper

Updated or fixed a deposit, or is there an erratum? Ask us to re-run the metrics. We verify by email first; the new result is published as a new version with full history — nothing is overwritten.

Reason for the rerun

We email you a confirmation link first. The rerun is an objective re-measurement — it cannot change the verdict in your favour, only ask us to look again.

Provenance — full disclosure

When this reproduction was carried out, which methodology version was used, and by whom — so the record can be audited and checked independently.

Reproduced
2026-06-14
Rubric version
v1.0
Assessed by
🤖 AI curator · claude (ai-curator room) · v1.0 · run #1 2026-06-15
no human curator yet
Last updated
2026-09-19

Provisional, curator- or AI-assessed, and independently checkable. A reproduction outcome states what one attempt could reproduce — not a judgement of the authors.

Deep full-text extraction

Model: sonnet
Founding hypothesis

The study tests whether integrating machine learning with multiomic data (transcriptomic, epigenetic, and mutational) on lactylation-related genes (LRGs) can enable molecular subtyping and accurate prognosis prediction in lung adenocarcinoma (LUAD).

Core claims
  • Ten multiomics clustering algorithms identify two distinct lactylation cancer subtypes (CS1 and CS2) in LUAD finding
  • CS1 is associated with significantly better overall survival than CS2, validated in independent cohorts finding
  • Nine hub LRGs (HNRNPC, PPIA, BZW1, GAPDH, H2AFZ, RAN, KIF2C, RACGAP1, WBP11) can be used to build a machine learning-based prognostic model method
  • A random survival forest (RSF) model based on the nine hub LRGs stratifies patients into high- and low-risk groups with strong prognostic value method
  • High-risk patients show more advanced TNM/clinical stage, higher recurrence/metastasis, and lower 1-, 3-, 5-year overall survival finding
  • Most oncogenic mutations occur in the high-risk group finding
  • Low-risk patients exhibit higher immune activity, suggesting better response to immunotherapy finding
  • Low-risk patients show greater predicted sensitivity to chemotherapeutics via oncoPredict analysis finding
Experimental setups
Assay System Perturbation Readout Platform
multiomics ensemble clustering (mRNA/lncRNA transcriptome, DNA methylation, somatic mutation) TCGA LUAD cohort none lactylation cancer subtype assignment (CS1/CS2) MOVICS R package (CIMLR, ConsensusClustering, SNF, iClusterBayes, PINSPlus, moCluster, NEMO, IntNMF, COCA, LRA)
closest template prediction / survival subtyping validation GSE13213 LUAD cohort (117 patients) none subtype classification and overall survival
closest template prediction / survival subtyping validation GSE31210 LUAD cohort (226 patients) none subtype classification and overall survival
ssGSEA pathway/gene set enrichment TCGA LUAD cohort (CS1 vs CS2) none differential pathway enrichment scores (BioCarta, HALLMARK, KEGG, Reactome) GSVA R package / MSigDB
immune cell infiltration profiling TCGA LUAD cohort (CS1/CS2 and high/low-risk groups) none immune cell infiltration scores IOBR (MCPcounter, EPIC, xCell, CIBERSORT, IPS, quanTIseq, ESTIMATE, TIMER)
univariate/multivariate Cox regression TCGA, GSE31210, GSE13213 LUAD cohorts none hazard ratio and p-value per LRG
machine learning survival modeling (79 algorithm combinations) TCGA LUAD training cohort none C-index, risk score stratification CoxBoost, stepwise Cox, Lasso, Ridge, Elastic Net, survival SVM, GBM, supervised PCA, partial least Cox, random survival forest
drug sensitivity prediction TCGA LUAD cohort (high vs low risk) none predicted IC50 for chemotherapeutics oncoPredict R package / Genomics of Drug Sensitivity in Cancer database
Key results
  • CS1 subtype associated with significantly better overall survival than CS2 in TCGA cohort p<0.001
  • CS1 vs CS2 survival difference validated in GSE13213 p=0.002 (n=117)
  • CS1 vs CS2 survival difference validated in GSE31210 p<0.001 (n=226)
  • RSF algorithm achieved the highest mean C-index among 79 combinations and was selected for the final model
  • ROC AUCs for 1-, 3-, 5-year OS prediction across cohorts TCGA: 0.956/0.975/0.954; GSE13213: 0.730/0.683/0.668; GSE31210: 0.833/0.793/0.783
  • High-risk group enriched for stage T3+4, N1+2+3, M1, III+IV disease, higher recurrence/metastasis, shorter survival p<0.05
  • Most oncogenic mutations detected in high-risk group
  • Immune activity significantly elevated in low-risk patients; low-risk patients show lower IC50 (greater chemosensitivity)
Key statistics
  • pvalue p<0.001 (OS difference between CS1 and CS2 in TCGA LUAD cohort)
  • pvalue p=0.002 (CS1 vs CS2 survival validation in GSE13213)
  • pvalue p<0.001 (CS1 vs CS2 survival validation in GSE31210)
  • other AUC 0.956, 0.975, 0.954 (1-, 3-, 5-year OS ROC AUC in TCGA training set)
  • other AUC 0.730, 0.683, 0.668 (1-, 3-, 5-year OS ROC AUC in GSE13213)
  • other AUC 0.833, 0.793, 0.783 (1-, 3-, 5-year OS ROC AUC in GSE31210)
  • count 336 total LRGs; 28 signature LRGs; 9 hub LRGs (gene set sizes used across subtyping and model construction)
  • pvalue p<0.05 (chi-squared/Wilcoxon test for TNM staging, clinical stage, and recurrence/metastasis differences between risk groups)

Statistical methods review

Model: sonnet

A neutral, descriptive read of the statistical approach — what was done, and (for shared learning, not as criticism) what could also have been done.

The paper applied a consensus multiomics clustering approach—combining mRNA, lncRNA, DNA methylation, and somatic mutation data from TCGA LUAD—across 10 clustering algorithms to derive two lactylation cancer subtypes, subsequently validated in two independent GEO cohorts. Hub LRGs selected by univariate Cox regression across three cohorts were used to build 79 machine-learning algorithm combinations, with the Random Survival Forest model chosen by highest mean C-index. The prognostic model stratified patients into high- and low-risk groups evaluated by Kaplan–Meier survival analysis, time-dependent ROC/AUC, and univariate/multivariate Cox regression; between-group differences in clinical and molecular features were tested with Wilcoxon rank-sum, Student's t-test, ANOVA, Kruskal–Wallis, chi-squared, and Fisher's exact tests, with a blanket significance threshold of p < 0.05 and FDR < 0.25 applied specifically to GSEA pathway enrichment.

Replicationbiological Sample sizeGSE13213: 117 patients; GSE31210: 226 patients; TCGA training-cohort size not explicitly stated in the text GroupsCS1 vs CS2 (multiomics consensus subtypes); high-risk vs low-risk (RSF prognostic model score) Pairingunpaired Randomization/blindingnot stated Dispersionnone Exact p-valuesno Effect sizesyes Confidence intervalsno Multiplicity correctionBenjamini-Hochberg FDR for GSEA/clusterProfiler only; no correction stated for other families of tests
Statistical tests used
Test Applied to n Assumptions
Kaplan–Meier survival analysis (log-rank test implied by reported p-values) Overall survival comparison of CS1 vs CS2 subtypes and high- vs low-risk groups in TCGA, GSE13213, and GSE31210 cohorts GSE13213: 117; GSE31210: 226; TCGA training-set n not stated in text not stated
Wilcoxon rank-sum test Non-normally distributed variable comparisons between two groups (e.g., TNM staging between high- and low-risk groups) not stated
Unpaired Student's t-test Normally distributed variable comparisons between two groups not stated
One-way ANOVA Parametric variable comparisons involving more than two groups not stated
Kruskal–Wallis test Non-parametric variable comparisons involving more than two groups not stated
Two-sided Fisher's exact test Contingency table analyses not stated
Chi-squared test TNM staging, clinical staging, recurrence/metastasis rate, and OS comparisons between high- and low-risk groups not stated
Univariate Cox proportional-hazards regression Screening each of 28 LRGs across TCGA, GSE31210, and GSE13213 to identify hub LRGs (p < 0.05, HR > 1 for 'bad'; HR < 1 for 'good') GSE13213: 117; GSE31210: 226; TCGA n not stated not stated
Multivariate Cox proportional-hazards regression Independent prognostic evaluation of risk score vs age, sex, and TNM/clinical stages in TCGA LUAD not stated
Time-dependent ROC / AUC (survivalROC package) 1-, 3-, and 5-year OS discrimination in TCGA, GSE13213, and GSE31210 GSE13213: 117; GSE31210: 226; TCGA n not stated not stated
GSEA with FDR correction (clusterProfiler) GO, KEGG, and HALLMARK pathway enrichment between high- and low-risk groups; criterion FDR < 0.25 and |NES| > 1 not stated
ssGSEA (via GSVA package) Immune cell infiltration scoring compared across CS1/CS2 subtypes and risk groups na
Cox regression-based gene filtering (MOVICS 'getElites', method = 'cox') Initial screening of 336 LRGs to select prognostic genes at p < 0.01 per data dimension across mRNA, lncRNA, methylation, and mutation layers not stated
Approaches that could also have been used
  • The best-performing algorithm among 79 combinations was selected by highest mean C-index computed on the TCGA training cohort itself
    Could also: Internal cross-validation (e.g., repeated k-fold or bootstrap resampling) or a held-out temporal/geographic split could also be used to estimate and compare C-indices before final model selection — Selecting a model by its training-set C-index can yield optimistic performance estimates; cross-validated or bootstrap-corrected C-indices provide a less upwardly biased comparison across algorithm candidates, and the magnitude of any optimism can then be reported alongside the final estimate
  • Twenty-eight LRGs were screened for hub status via separate univariate Cox regressions in each of three cohorts at p < 0.05, and the intersection across cohorts was taken as the final gene set
    Could also: A penalized Cox regression (Lasso or elastic net) applied jointly to all candidate genes, or a fixed-effects meta-analysis pooling log-HR estimates across the three cohorts, could also be used for gene selection — Running many separate univariate tests without multiplicity correction inflates the false-positive rate; penalized regression handles inter-gene correlation directly, and meta-analytic pooling quantifies between-cohort heterogeneity rather than requiring perfect replication as an implicit filter
  • Overall survival comparisons between groups were presented as Kaplan–Meier curves with p-values only
    Could also: Reporting median survival with 95% confidence intervals and hazard ratios with CIs alongside the KM plots could also be done — Log-rank p-values convey whether a difference exists but not its magnitude or clinical relevance; median survival times with CIs and HR with CIs give readers a more complete picture of the size and precision of the observed survival difference
  • AUC values for 1-, 3-, and 5-year OS prediction were reported as point estimates without confidence intervals across three cohorts
    Could also: Bootstrap-derived 95% CIs around each AUC, or the DeLong method for paired AUC comparisons, could also be reported — AUC point estimates vary with sample size and cohort composition; CIs communicate the precision of the discrimination estimate and allow readers to judge whether differences across time points or between cohorts reflect meaningful variation or sampling variability
  • Immune cell infiltration was estimated using eight separate deconvolution algorithms (MCPcounter, EPIC, xCell, CIBERSORT, IPS, quanTIseq, ESTIMATE, TIMER) and results were reported in aggregate
    Could also: Pre-specifying one algorithm as the primary analysis and treating the others as sensitivity checks, or deriving a consensus score across algorithms, could also structure the inference — Running eight parallel deconvolution comparisons generates an implicit multiple-comparison burden; a pre-specified primary method with pre-registered sensitivity analyses makes the confirmatory inference explicit and reduces the risk that a result prominent in the narrative was selected post-hoc from a favorable algorithm
  • Continuous variables were assigned to either parametric (t-test) or non-parametric (Wilcoxon) branches based on distributional assumptions, but the method used to assess normality was not stated
    Could also: Formally reporting a normality assessment (e.g., Shapiro-Wilk test or Q-Q plot inspection) or applying the Wilcoxon rank-sum test uniformly across all two-group comparisons could also be used — Describing the criterion for the parametric/non-parametric split allows readers to evaluate whether the test selection was appropriate; alternatively, using a single non-parametric approach consistently removes the need for a normality decision and is often recommended when sample sizes are small or heterogeneous across groups
Software: R 4.2.2 · MOVICS (R package) · maftools (R package) · GSVA (R package) · pheatmap (R package) · survminer (R package) · survivalROC (R package) · IOBR (R package) · clusterProfiler (R package) · limma (R package) · oncoPredict (R package)

Citation network

Where this publication sits in the reproducibility-weighted citation graph — what it is built on, and what is built on it. Citation data from OpenAlex.

Authors · 2
Citations
8
Impact: low
Foundation confidence
None of its references are in our reproducibility record yet — its foundation cannot be assessed.
Topics

No assessed neighbours yet — the network grows as more papers are assessed.

Data lineage

The datasets this paper uses (text-mined from the full text via Europe PMC), and which other assessed papers stand on the same data. A shared dataset is a factual link — not a judgement.

GSE31210 GEO in Discussion (http://purl.org/orb/Discussion)
also used by 3 papers:
GSE13213 GEO in Discussion (http://purl.org/orb/Discussion)
no other assessed paper uses this yet

What was reproduced

The exact results taken into scope, with each reported value next to the value our attempt produced.

Scope — pmid-39856156

Paper: Hua M, Li T. Multiomic machine learning on lactylation for molecular typing and prognosis of lung adenocarcinoma. Sci Rep 2025. DOI 10.1038/s41598-025-87419-4 · PMID 39856156 · PMCID PMC11760357. Code: https://github.com/ltzxxz/LUAD_lactylation @ commit 2bbb5b53c005395e9da45d4f1f54b53283161a8d (R scripts 01–17; authors' own code). Data: TCGA-LUAD (UCSC-Xena, public), GEO GSE31210 + GSE13213 (public).

Data availability reality

The repo README states "Any data produced during this pipeline is available upon request." The intermediate objects (cleandata.RDATA, markergene.txt, RS.RDATA, 筛选基因表达临床.RDATA) and the TCGA preprocessing inputs (TCGA.txt, …LNCRNA.TXT, …varscan.gz, …methylation…txt, clinicaldata.txt) are NOT shipped. But the raw sources are public: TCGA-LUAD on UCSC-Xena and the two GEO series. The final model is fully specified in code + paper (9 hub genes named, algorithm = RSF, exact rfsrc params), so it is reconstructable.

IN SCOPE (attempted) — pipeline-derived, low-hanging (80/20)

Result Paper loc Pipeline Status
9-gene RSF risk model → GSE31210 1/3/5-yr OS AUC = 0.833/0.793/0.783 Fig 8B 08_Machine_learning.R block 1-1 RSF: per-cohort scale()rfsrc(ntree=1000,nodesize=5,logrank,seed=123)predict$predictedtimeROC reproducing
9-gene RSF risk model → GSE13213 1/3/5-yr OS AUC = 0.730/0.683/0.668 Fig 8A same reproducing
RSF best-algorithm / C-index ranking + TCGA risk-score concordance Fig 7B/E same model, Cindex per cohort supportive

The 9 hub genes (paper text): HNRNPC, PPIA, BZW1, GAPDH, H2AFZ, RAN, KIF2C, RACGAP1, WBP11.

Faithfulness note

Validation cohorts (GSE31210, GSE13213) are mRNA-only → fully reconstructable from GEO. The TCGA training set is the single approximation: the paper trained on multiomic-complete TCGA samples (intersection of mRNA/lncRNA/methylation/ mutation/clinical — that exact sample list is not shipped). We train on all TCGA-LUAD primary tumours with OS. Therefore the RSF (even at seed=123) is not byte-identical; AUCs are expected within-tolerance, not exact. This is reported honestly, not as ground truth.

OUT OF SCOPE (not attempted) — needs non-shipped intermediates / heavy/manual

  • MOVICS 10-algorithm multiomic clustering → 2 cancer subtypes (CS1/CS2) (02_Clustering.R, Fig 1–3): requires cleandata.RDATA (4-omic TCGA matrices) not shipped; reconstructing all four omics + the consensus of 10 clusterers is the hard >20%. Skipped.
  • GSVA/ssGSEA pathway scoring, IOBR immune deconvolution, stemness/TMB/MSI, microenvironment, chemokine, oncoPredict drug-resistance, maftools mutation panels (scripts 03–06, 10–17): downstream of the non-shipped subtype/score objects and/or require the multiomic intermediates. Skipped.
  • The candidate-gene (markergene.txt) derivation and univariate-Cox 3-cohort intersection (07_Gene_selection.R): markergene.txt not shipped; we instead use the paper's stated final 9 genes directly (faithful to the reported model).

Possible-fabrication watch

None flagged a priori — the 9 genes and RSF params are concretely specified and the validation data is public, so the AUC claims are checkable. The reproduced AUCs (this run) are the evidence a human can compare against Fig 8.

Figures / tables: Fig 8BFig 8AFig 7Fig 7BFig 8Fig 3B
GSE31210_AUC1
Reported
0.833
Reproduced
0.810
within tolerance
GSE31210_AUC3
Reported
0.793
Reproduced
0.754
within tolerance
GSE31210_AUC5
Reported
0.783
Reproduced
0.756
within tolerance
GSE13213_AUC1
Reported
0.730
Reproduced
0.794
within tolerance
GSE13213_AUC3
Reported
0.683
Reproduced
0.725
within tolerance
GSE13213_AUC5
Reported
0.668
Reproduced
0.720
within tolerance
hub_genes_9
Reported
9 genes HNRNPC,PPIA,BZW1,GAPDH,H2AFZ,RAN,KIF2C,RACGAP1,WBP11
Reproduced
same 9 genes, all present in TCGA + both GEO platforms
exact
best_algo_RSF
Reported
RSF best mean C-index
Reproduced
RSF C-index 0.736 (GSE31210) / 0.699 (GSE13213) out-of-sample
partial
GSE31210_KM
Reported
p<0.001
Reproduced
median-split logrank p=2.6e-04
within tolerance

Assessments & scoring basis

Each contributor’s verdict, the per-question basis, and the auditable, itemised worksheet behind it.

🤖 AI curator · claude (ai-curator room) · v1.0 L1 83/100

An automated assessment. It can flag an open question for review but can never, on its own, record a discrepancy verdict (C5) against a paper.

🟡1. Data identity
🟢2. Endpoint comparability
🟡3. Location of the main deviation
🟡4. Cause of the deviation
🟢5. Derivability / plausibility
🟡6. Severity of the deviation
🟢7. Core claim
🟡8. Severity of the miss (overall human judgment)
Scoring basis — itemised

Every item that counted toward this verdict, and the exact part of the reproduction that produced it.

Supporting (toward a concern)
Content-critical question only partially held
+2 pts
From: Q8 · Severity of the miss (overall human judgment) 🟡
Minor / cosmetic deviation
+1 pts
From: Q3 · Location of the main deviation 🟡
Minor / cosmetic deviation
+1 pts
From: Q4 · Cause of the deviation 🟡
Minor / cosmetic deviation
+1 pts
From: Q6 · Severity of the deviation 🟡
Concordant (toward reproduced)
Code + data deposited & functional
-2 pts
From: Data & code availability Available & functional
Total score +3

The paper's headline prognostic claim — a 9-gene lactylation Random Survival Forest validated on public GEO cohorts — reproduces well: all 6 Fig 8 AUCs land within ±0.065 (same direction, GSE13213 even higher), all 9 hub genes recover, RSF gives valid out-of-sample separation (p=2.6e-04). The residual deviations sit on the input/training side and are our-method-driven: the authors' exact TCGA multiomic-complete training subset and normalization were not deposited, so we approximated with 515 TCGA-LUAD samples. This is a data-availability/cohort-definition limitation, not a fabrication signal — values are derivable and the core conclusion holds. Graded yellow overall because of the explainable training-set deviations and because only the prognostic half was in scope.

🤝
Reproduced automatically — and fairly

Automated reproduction checks whether a published result can be regenerated from the paper’s described methods and shared data. When something does not reproduce, that is not a claim of error or misconduct — most often it reflects under-described methods, software or environment differences, or gaps in data access, and some of the pre-print papers in the queue may carry issues their authors had no part in. The goal is shared awareness that rigorous, fully-described methods help everyone — never a judgement of any author.

Are you an author? We would genuinely like to hear from you — to clarify the record, add data or code, re-run the pipeline after an accession update, and publish your response right next to the assessment. Everything here is open and auditable.

🚩 Report an error in this record

Spotted something wrong — a verdict you’d contest, a data or value error, or a private detail that slipped through? Tell us, with a short justification. Authors and readers are equally welcome to write in; we review every report.

Prefer email, or the form below not working? Contact us at [email protected].

Reproduction footprint

claude-opus-4-8

Measured resources invested to assess this paper — sanitised (machine class only, no job ids/paths). Compute = HPC accounting (SLURM); tokens = the AI agent's session.

151.6 k
tokens (I/O) · 9.1 M incl. cache
16 min
runtime · 0.03 CPU-h
2.5 GB
peak RAM
3 (2 failed)
HPC jobs
hummel
machine