Corpus 1,272 assessed · 1,173 scored · 643 reproduced ≥75 · 168 flagged ·∅ 74.1/100
← New search

A hybrid empirical and parametric approach for managing ecosystem complexity: Water quality in Lake Geneva under nonstationary futures.

Proc Natl Acad Sci U S A · 2022
L1 69/100 3/4
Why this verdict

The main results reproduced, with only marginal, non-material deviations.

Reproduced on the brainbox compute brainarbeit.com
✓ What held up
  • Same input data as the authors
  • Reported values were directly comparable
What did not (or only partly)
  • 🟡A deviation arose in the data or preprocessing
  • 🟡A deviation was attributed to the published material
  • 🟡Reported values were not (fully) derivable from the shared data
  • 🟡The deviation was non-trivial in magnitude
  • 🟡The central claim did not (fully) hold under reproduction
  • 🟡Overall, the reproduction showed a material discrepancy
How its reproducibility compares
69/100
Reproducibility score
0.3 SD below mean
vs. all fields · 1173 studies
🎯 Scores higher than 35% of all assessed papers rank 745 of 1173 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

REPRODUCED (1:1 on the headline result). PNAS 2022 Lake Geneva hybrid empirical(EDM)+parametric(Simstrat) DO model; authors' own R repo SugiharaLab/Geneva_Hybrid @91739e3 driven by legacy rEDM 0.7.x on shipped monthly limnological data (Zenodo 6587597 == repo snapshot, 545 monthly records). Environment rebuilt on «our HPC»: conda R 4.1.3 + dplyr 1.0.10 + rEDM 0.7.3 (compiled from CRAN archive); ran after a minor Linux path-case fix (DATA/LIB symlinks). The two headline numbers reproduce essentially exactly: hybrid DOB hindcast rho=0.898 (paper 0.89) and MAE=0.951 mg/L (paper 0.94); parametric baseline correctly far lower (0.651). CCM tables (C4/C5), Fig1D mEDM skill (C6), and greedy-vs-baseline models (C7/C8) all reproduce -- and the shipped precomputed mEDM caches recompute BIT-FOR-BIT from code+data (max diff ~1e-16), i.e. no fabrication in the deposited intermediates. The TP x warming hypoxia scenario grid (C10, 138 hybrid runs) reproduces the full monotonic structure with corners within ~1.5-8 percentage points of the figure-read values. NOT separately quantified: C9 (S-map coefficient series, ran but not isolated) and C11 (seasonal depletion-rate metric, scenario data saved). OUT OF SCOPE (not attempted): the Simstrat hydrodynamic model itself (parametric/Windows-binary; its outputs are shipped and consumed as fixed EDM inputs) and field/wet-lab monitoring (external data collection).

💻 Code ↗ 🗄 Data: 10.5281/zenodo.6587597

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 69
    assessed: 2026-06-21 ⛓ 3bc02c749203
✎ 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
2026-06-21
Rubric version
v1.0
Assessed by
🤖 AI curator · claude (ai-curator room) · v1.0 · run #1 2026-06-19
no human curator yet
Last updated
2026-08-05

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 paper asks whether the major interdependencies among lake physics, biogeochemistry, and ecology driving deep-water dissolved oxygen (DOB) in Lake Geneva can be resolved without building a model too complex to reliably fit or interpret, particularly given confounded, nonstationary effects of reoligotrophication and climate change.

Core claims
  • A hybrid model combining empirical dynamic modeling (EDM/S-map) for biogeochemical sink terms with the equation-based Simstrat hydrodynamic model for physical source terms produces substantially better historical DOB forecasts than conventional parametric models. method
  • The effects of total phosphorus (TP) on chlorophyll (CHL) and of CHL on DOB are state-dependent (nonlinear) rather than fixed, changing systematically as reoligotrophication progresses. finding
  • A moderate 3°C increase in air temperature would impact Lake Geneva water quality on the same order of magnitude as the eutrophication that occurred over the previous century. finding
  • Convergent cross-mapping (CCM) identifies air temperature, lake temperature, thermal structure, CHL, and phosphorus as causal drivers of DOB. finding
  • Nonlinear (state-dependent) S-map models with θ > 0 substantially outperform the linear multivariate autoregressive (MAR) equivalent (θ = 0) in predicting DOB. finding
  • Multiple modes of management intervention (both nutrient control and climate mitigation) are likely necessary to achieve a healthy lake. finding
  • S-map regression provides a general empirical dynamic modeling framework for quantifying time-varying (Jacobian-like) interaction strengths among ecosystem variables. method
Experimental setups
Assay System Perturbation Readout Platform
Convergent cross-mapping (CCM) causality analysis Lake Geneva long-term limnological time series none (observational) causal coupling among DOB and candidate drivers (air temp, lake temp, thermal structure, CHL, phosphorus) Empirical Dynamic Modeling (EDM)
Multivariate S-map (nonlinear state-space) regression Lake Geneva time series (hmix, Tsurf, Tatm, Q, CHL, TPsurf, TPlake, DOB) none (sequential variable addition to embedding) forecast skill (Pearson correlation between observed and predicted DOB) as function of nonlinearity parameter θ EDM / S-map
S-map partial derivative (Jacobian coefficient) estimation Lake Geneva time series none (state-dependent analysis across TP levels) ∂CHL/∂TP and ∂DOB/∂CHL as functions of TPlake state EDM / S-map
1D hydrodynamic modeling Lake Geneva water column none / atmospheric forcing input daily thermocline depth prediction accuracy Simstrat model
Hybrid empirical–parametric scenario modeling Lake Geneva (simulated 6-month stratified periods) fixed-background TP concentrations (reoligotrophication states) and air-temperature increase scenarios DOB depletion rate over stratified season Simstrat + EDM hybrid model
Sediment core analysis Lake Geneva sediments none (historical reconstruction) historical hypoxia status relative to TP levels
Key results
  • Simstrat reproduces daily thermocline depth across four decades with low error nrmse = 6.7% (rmse = 21 m)
  • Adding biogeochemical variables (CHL, TP surf, TP lake) to the physical embedding sequentially improves multivariate DOB forecast skill
  • Nonlinear S-map models (θ > 0) predict DOB much better than the linear MAR model (θ = 0)
  • CHL showed essentially no response to TP when TP lake was high (1980s); phosphorus limitation of CHL only emerged once TP lake fell below a threshold threshold ~40 µg·L⁻¹ (near-zero effect at TP lake > 60 µg·L⁻¹)
  • DOB was insensitive to CHL during the eutrophic period but developed a consistent negative response to CHL as reoligotrophication proceeded threshold ~50 µg·L⁻¹ TP lake
  • Lake Geneva sediment cores show the lake was not hypoxic prior to 1945 TP < 10 µg·L⁻¹
  • Hybrid model scenarios show DOB depletion rate is high at high background TP and decreases with lower TP, with warming scenarios producing effects comparable to historical eutrophication
Key statistics
  • other nrmse = 6.7%, rmse = 21 m (Simstrat thermocline depth prediction accuracy over four decades)
  • other TP lake < 40 µg·L⁻¹ (threshold below which CHL begins responding to phosphorus limitation)
  • other TP lake < 50 µg·L⁻¹ (threshold below which DOB shows consistent negative response to CHL)
  • other TP < 10 µg·L⁻¹ (phosphorus level prior to 1945 when lake was not hypoxic (sediment core evidence))
  • other TP lake > 60 µg·L⁻¹ (1980s phosphorus level at which mean annual effect of TP on CHL was essentially zero)
  • other DOB_init = 7.5 and 6 mg·L⁻¹ (initial dissolved oxygen conditions used in hybrid model stratified-period scenario experiments)
  • other 3°C air temperature increase (climate scenario whose predicted water-quality impact is comparable to 20th-century eutrophication)
  • other P<36 µg·L⁻¹ (phosphorus level below which initial TP-reduction actions began to affect algal biomass and DOB)

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 uses a two-stage empirical/parametric hybrid modeling approach rather than classical hypothesis testing: first, Empirical Dynamic Modeling (EDM), including convergent cross-mapping (CCM) for causal inference and multivariate S-map regression for state-dependent nonlinear prediction, is used to identify and quantify time-varying relationships among lake variables (TP, CHL, DO_B, physical drivers); second, this empirical model is combined with a deterministic, equation-based hydrodynamic model (Simstrat) to forecast dissolved oxygen decades into the future under different climate/nutrient scenarios. Model performance is reported primarily as forecast skill (Pearson correlation between observed and predicted values) and error metrics (nrmse, rmse) rather than through group-comparison significance tests.

Replicationunclear Sample sizeTime spans are described qualitatively (e.g., 'four decades' of daily data for Simstrat validation) but explicit sample sizes for the causality/S-map analyses, or any power analysis, are not given in the text provided. GroupsNot a group-comparison design; the study compares model forecast skill across variable embeddings and across TP/temperature scenarios, and compares hybrid vs. parametric-only model predictions over time. Pairingna Randomization/blindingna Dispersionunclear Exact p-valuesno Confidence intervalsno
Statistical tests used
Test Applied to n Assumptions
Convergent cross-mapping (CCM), a nonlinear time-series causality detection method Causal driver identification for DO_B (Fig. 1 B and C, SI Appendix Fig. S2, Table S1) not stated
Multivariate S-map (sequential locally weighted global linear map) regression, tuned by nonlinearity parameter θ Prediction of DO_B and CHL from embeddings of driver variables (Fig. 1D, Fig. 2, SI Appendix Figs. S3, S5, S6) not stated
Comparison of linear MAR model (S-map with θ=0) versus nonlinear S-map (θ>0) forecast skill Fig. 1D, evaluating predictability of DO_B across embeddings not stated
Forecast skill quantified as Pearson correlation between observed and predicted values Multivariate EDM/S-map predictions of DO_B (Fig. 1D) and hybrid model validation against historical data na
Deterministic hydrodynamic model (Simstrat) validated with normalized RMSE and RMSE Prediction of daily thermocline depth over four decades four decades of daily thermocline depth data (exact n not given) not stated
Approaches that could also have been used
  • Causal relationships among lake variables were assessed with convergent cross-mapping (CCM), a nonlinear, state-space-based causality detection method.
    Could also: Granger causality testing (or its nonlinear extensions) is a standard alternative for inferring predictive causal relationships in time series. — Granger causality is widely used and well understood for linear or weakly nonlinear systems, and comparing its results with CCM could help characterize how much of the detected causality depends on nonlinear state-dependence versus simpler linear predictive relationships.
  • State-dependent relationships (e.g., effect of TP on CHL, CHL on DO_B) were quantified via S-map regression coefficients without reported confidence intervals or uncertainty bounds.
    Could also: Bootstrap resampling (e.g., block bootstrap for time series) could also be used to generate confidence intervals around the time-varying S-map coefficients. — This would let readers gauge the sampling uncertainty of the estimated Jacobian-like coefficients shown in Fig. 2, complementing the point estimates already presented.
  • Model predictability was reported as forecast skill (Pearson correlation between observed and predicted DO_B) across different embeddings and θ values.
    Could also: Reporting additional accuracy metrics such as RMSE, MAE, or a formal cross-validated log-likelihood alongside correlation could also be used. — Correlation-based skill can be insensitive to systematic bias or scale errors, so pairing it with an absolute error metric can give a fuller picture of predictive accuracy, similar to how nrmse/rmse were already used for the Simstrat model.
  • The nonlinearity tuning parameter θ was varied to compare linear (θ=0, equivalent to a MAR model) versus nonlinear S-map fits, with the best θ apparently chosen by inspection of the forecast-skill curve (Fig. 1D).
    Could also: A formal leave-one-out or blocked time-series cross-validation procedure with a defined selection criterion (e.g., maximum out-of-sample skill with reported variability across folds) could also be used to select θ. — This would make the θ selection procedure fully reproducible and would allow quantification of how sensitive the chosen embedding is to the particular data split used.
  • The hybrid empirical-parametric model's improved performance is described qualitatively as leading to 'substantially better forecasts' relative to the parametric-only approach.
    Could also: A formal model comparison framework, such as comparing information criteria (AIC/BIC where applicable) or a paired test on out-of-sample forecast errors between the hybrid and parametric-only models, could also be used. — This would provide a quantitative, statistically grounded comparison of the two modeling approaches rather than a descriptive comparison of forecast plots.
  • Scenario projections (different TP and air-temperature combinations) are presented as point trajectories of DO_B depletion (Fig. 3) without stated uncertainty ranges.
    Could also: Ensemble or Monte Carlo propagation of parameter and initial-condition uncertainty through the hybrid model could also be used to generate prediction intervals around the scenario trajectories. — This would communicate the range of plausible outcomes under each climate/nutrient scenario, which can be useful for management decisions that depend on the confidence in projected hypoxia risk.
Software: Simstrat (1D hydrodynamic model)

What was reproduced

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

Scope — pmid-35733249

Paper: Deyle ER, Bouffard D, Frossard V, Schwefel R, Melack J, Sugihara G (2022). "A hybrid empirical and parametric approach for managing ecosystem complexity: Water quality in Lake Geneva under nonstationary futures." PNAS 119(26):e2102466119. PMID 35733249 · PMCID PMC9245694 · DOI 10.1073/pnas.2102466119

Code: https://github.com/SugiharaLab/Geneva_Hybrid @ commit 91739e31e3ea448cebea97d77e23e90fab4de693 (resolved 2026-06-19, latest HEAD). Authors' own repo (P16 not relevant). 100% R. Data: Zenodo 10.5281/zenodo.6587597 + data also shipped inside the repo.

Nature of the study

Empirical Dynamic Modeling (EDM: simplex, S-map, CCM, multivariate block_lnlp, multiview via the legacy rEDM 0.7.x API — repo loads library("rEDM", lib.loc="./LIB"), "the archived version of rEDM") applied to ~monthly limnological time series of Lake Geneva (CARRTEL/SOERE monitoring), combined with a parametric hydrodynamic model (Simstrat) to build a "hybrid" forecast of deep-water dissolved oxygen (DOB). Not a bioinformatics pipeline — but the reproducible computational core is the EDM analysis run with a published third-party/legacy R package on the paper's own data, which the brief explicitly counts as equally valid.

IN SCOPE — pipeline-derived (EDM) results we attempt to reproduce

Pipeline = rEDM 0.7.x driven by the repo's Rmd notebooks (PNAS_SI_notebook.Rmd, analysis_chl.Rmd, analysis_po4_epi.Rmd).

id result where in paper how computed
C1 Hybrid DOB hindcast skill ρ = 0.89 (Obs vs Hybrid, 1981–2017) Fig 4A, text hybrid_model_v1() + EDM-simulated chl/PO4 + Simstrat physics; cor() in SI notebook L1200
C2 Hybrid DOB hindcast MAE = 0.94 mg/L Fig 4A, text mean(abs(Obs-Hybrid)) SI L1202
C3 Parametric (Simstrat/"Robert") DOB skill — comparison baseline Fig 4A cor(Obs, Parametric) SI L1201; directly checkable from two shipped CSVs
C4 Table S1: CCM cross-map skill of DO_delta drivers (ρ, sorted) Table S1 ccm() per driver, E* by max ρ, tp=-floor(E*/2)
C5 Fig S4 / Table S5: CCM skill from chl and PO4_epi Fig S4 panels ccm() as C4
C6 Fig 1D: multivariate EDM forecast skill (ρ vs θ) for DOB, sequential variable addition Fig 1D do_mEDM_models() over 5 embeddings, S-map θ sweep
C7 mEDM "greedy" model for chl (embedding, θ, ρ, mae, rmse) vs baselines (univar/seasonal/multiview) text/SI, supports hybrid do_mEDM_greedy(), multiview(), s_map()
C8 mEDM "greedy" model for PO4_epi vs baselines SI as C7
C9 Fig 2 state-dependent S-map coefficients: ∂CHL/∂TP sign change near TP≈40 µg/L; ∂DOB/∂CHL threshold Fig 2, text block_lnlp(..., save_smap_coefficients=T, theta=6)
C10 Scenario hypoxia: % time DOB<4 mg/L = 55/20/55/85 % across TP×ΔT scenarios Fig 4B/C, text iterative hybrid_model_v1() over 45 TP × 3 ΔT (heavier)
C11 Seasonal deep-O2 depletion range 0–0.017 mg/L/day from EDM scenarios text derived from scenario forecasts

Priority / quick floor: C3 (CSV-only), C1, C2, C4, C6 are the low-hanging, clearly-specified outputs. C7–C11 are the harder "keep going" targets.

OUT OF SCOPE — not pipeline-derived (not attempted, or inputs-shipped only)

  • Simstrat hydrodynamic runs (thermocline NRMSE 6.7 %, rmse 21 m; deep-temp RMSE ~0.2 °C). This is the parametric half: a separate equation-based physics model (Fortran/Windows binary), not EDM. Its outputs (T_out.dat, scenario runs) are shipped in the repo, so the EDM/hybrid steps consume them as fixed inputs. We do NOT re-run Simstrat. (Could optionally cross-check shipped Simstrat temp vs observed, but that validates the parametric model, not the EDM pipeline.)
  • Field monitoring / wet-lab measurement (CARRTEL sampling of TP, SRP, CHL, DOB; MeteoSwiss met data; Rhône discharge). External data collection — out of scope.
  • **Robert/ob
Figures / tables: Fig 4ATableFig S4Fig 1DFig 2Fig 4B
C1
Reported
rho=0.89 (hybrid DOB historical hindcast, Obs vs Hybrid)
Reproduced
rho=0.898044
within tolerance
C2
Reported
MAE=0.94 mg/L (hybrid)
Reproduced
MAE=0.950557 mg/L
within tolerance
C3
Reported
parametric rho << hybrid (Obs vs Robert/Parametric)
Reproduced
rho=0.650869 (vs hybrid 0.898)
within tolerance
C4
Reported
Table S1 CCM DO_delta driver cross-map skills (ranked rho)
Reproduced
ranking regenerated; h_mix_model 0.574 / T_air 0.553 top, chl 0.135 bottom
partial
C5
Reported
Fig S4 / Table S5 CCM skill from chl and PO4_epi
Reproduced
both tables regenerated (16 drivers each)
partial
C6
Reported
Fig 1D mEDM forecast skill curves (5 sequential embeddings)
Reproduced
best embedding (phys+chl+SRP) max rho=0.612
within tolerance
C7
Reported
greedy chl mEDM beats baselines
Reproduced
greedy 0.569 > univar 0.500 > seasonal 0.487; multiview 0.612; recompute IDENTICAL to deposit (4.4e-16)
within tolerance
C8
Reported
greedy PO4_epi mEDM beats baselines
Reproduced
greedy 0.960 > univar 0.943 >> seasonal 0.415; multiview 0.963; recompute IDENTICAL to deposit (2.2e-16)
within tolerance
C9
Reported
Fig 2 S-map interaction coeffs (sign/threshold ~TP 40)
Reproduced
chunk recomputed; coefficient series not isolated
partial
C10
Reported
hypoxia %: 20/55/55/85 across TP x dT
Reproduced
loess-smoothed: 27.6/60.1 (0C, TP25/60); 56.5/78.2 (3C, TP25/60); within ~8 pts; full structure reproduced
partial
C11
Reported
seasonal depletion 0-0.017 mg/L/day
Reproduced
scenario forecasts saved; depletion-rate metric not isolated
partial

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 69/100

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)
🤝
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 support@doesitreproduce.com.

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.

408.5 k
tokens (I/O) · 33.5 M incl. cache
85 min
runtime
Per-job HPC accounting not captured for this run — the runtime shown is the reproduction’s measured wall-clock time.