Methodology & validation — no black box

Source-attributed. Every equation and parameter mapped to its published origin; each tagged literature / calibrated / assumed.
Out-of-sample. Predictive performance assessed on data the models did not see during calibration, benchmarked against simple baselines, and reported by region and year — not as a single flattering average. [Metrics inserted on publication]
Uncertainty carried, not hidden. Predictive intervals are produced and checked for calibration; uncertainty is propagated through the cascades.
Structural channels labelled honestly. Counterfactual channels (pollination-collapse, biosecurity-recovery) are validated for plausibility, not predictive accuracy, and we say so.
Independently reviewed and versioned. Every release is frozen and shipped with a validation delta. Reviewed by [REVIEWER / on review]

A note on what we publish

We publish our methods and state our limitations. Validation metrics and named references appear here only once they are published and cleared for public use — we would rather underclaim than overclaim to an audience that models for a living.

Request the technical summary →