Reproducibility data and code for provenance aware evaluation and land accounting bounds in agricultural systems analysis

Description

This dataset accompanies the manuscript “Provenance aware evaluation and land accounting bounds for agricultural systems analysis”. It contains archived derived inputs, documented source ancestry, precomputation extension protocols, executable Python analysis and independent validation code, complete result tables, and figure-generation code for three agricultural statistics case studies in Zambia, Malawi and Niger. The Zambia analysis traces source observations entering both current forecast-productivity outcomes and five-year reference labels. It preserves original frozen predictions, audits training/test source sharing, evaluates an ancestry-disjoint subset, and fits fixed logistic comparators under spatial-only and joint source/spatial grouped cross-validation. The Niger analysis computes conditional bounds on unique crop-footprint area and combined reported grain mass per unique footprint hectare, then identifies within-year district orderings that remain supported under unobserved crop overlap. The integrated extension evaluates later-year transfer using historical-only training, frozen 2007–2011 training and simultaneous source-group/temporal separation on the same 444 later-year observations. The Niger extension compares summed-area benchmark selection with interval-box minimax regret, and reports conditional choice changes and common-overlap sensitivity. An optional food-energy conversion is explicitly a conditional scenario. Malawi archived verification records document source closure without changing reported values. The package is executable offline once its pinned Python dependencies are installed. Every supplied derived input is linked by SHA-256 hash to its source archive member. The package reproduces the downstream analyses from these derived inputs; it does not claim complete reacquisition of provider raw workbooks, regeneration of upstream harmonization, or refitting of the original XGBoost adaptation. New logistic fits and scoring transforms are reproduced. Original protocol records are preserved separately from post-review extension rules. Restricted KBS experimental records and KBS-derived quantitative summaries are excluded. The cases are purposive. Results are descriptive and conditional on the documented targets, source links, evaluation designs and land-accounting assumptions. They do not establish causal crop responses, prospective early-warning performance or a feasible land-reallocation policy.

Steps to reproduce

1. Download Agricultural_Systems_Reproducibility_v2.zip and extract it to a new directory. Work from the extracted package root containing README.md, inputs, protocols, scripts and results. 2. Use Python 3.12. Create and activate an isolated environment if desired. Install the exact dependencies with: python -m pip install -r requirements.txt 3. Before rerunning analyses, verify the distributed package: python scripts/check_manifest.py 4. Run the complete integrated workflow: python scripts/run_all.py 5. Confirm PASS for the 54 base checks and 23 decision/temporal validation check groups. Reports are results/validation_report.json and extensions/decision_temporal/results/validation_extension.json. 6. The workflow regenerates all five manuscript figures and the graphical abstract. To regenerate only the figures, use: python scripts/make_figures.py 7. Inspect the root results folder for source ancestry, grouped evaluation and accounting bounds; inspect extensions/decision_temporal/results for all later-year predictions, training support, annual decisions, interval-box regret and sensitivity tables. Repeated predictions and district pairs are not independent observations. 8. Consult input_provenance.json and docs/SOURCE_REGISTER.json for the upstream evidence boundary and original source references. Original XGBoost predictions and upstream raw-data reacquisition are outside the rerun scope. 9. The original audit, first computational extension and later decision/temporal extension remain distinct. Their protocol files are not external preregistration. The initial Niger pilot preceded the full decision protocol; the within-year AUC diagnostic was added after inspection of the pooled temporal results. 10. Read the target and interpretation limits. Zambia uses forecast production per planted hectare and retrospective season-end climate; Niger decisions use a district-wise interval-box relaxation. Neither operational early warning, observed management gains nor independent external empirical validation is established.

Institutions

• Kocaeli Üniversitesi

  Kocaeli, İzmit

Categories

Licence