Changelog • leafwax

leafwax 0.4.0

Replaced ratio inversion with a likelihood-based Bayesian inversion over paired calibration draws and a caller-specified reconstruction prior.
Multi-row records jointly reweight shared calibration draws and return explicit diagnostics for prior influence, slope sign, numerical integration, mixture effective sample size, and saved-draw stability.
Zero and negative slope draws are retained. Posterior samples require an explicit seed; the 100-draw preview tier cannot return inferential results.
Reconstruction fails closed for calibration variants whose complete new-site predictor design is unavailable. Model ensembles require an explicit list of compatible models and have no default scientific composition.
Public slope values now use physical units (per mil wax per per mil precipitation). local_effective_slope() back-transforms the fitted standardized coefficient, and invert_d2H() converts physical slope overrides back to fitted-model units internally. This also puts detect_change() thresholds on a consistent physical scale.

Posterior data and integrity update.

Posteriors were re-fit in the 2026-07-28 chordal-distance calibration analysis (n = 1,128 calibration observations; Africa 142). All 14 model posteriors, the shipped 100-draw preview tier, and the spatial-model knot metadata were generated from this run.
Three interaction models changed their coefficient set to match the fitted Stan specifications: elevation_c4_interact_sp carries the beta_oipc_x_c4 interaction term (previously absent), while full and full_sp no longer carry beta_oipc_x_c4 (retaining the plant-functional-type interactions beta_oipc_x_grass, beta_oipc_x_shrub, beta_oipc_x_tree).
Complete posterior files are distributed in companion data release v3.0.0 (doi:10.5281/zenodo.21880665). download_model_data() verifies both file size and SHA-256 against the immutable release manifest before a file enters the package cache.

CRAN release: 2026-06-15

CRAN compatibility updates.

Updated DESCRIPTION formatting for CRAN: software/service names are quoted as 'Stan' and 'Zenodo', and the data-deposit reference now uses Bradley (2026) <doi:10.5281/zenodo.20085465>.
Replaced \dontrun{} examples with executable \donttest{} or unwrapped examples, with temporary cache directories and quiet output where examples touch downloads or cache state.
Added the missing return-value documentation for print.leafwax_posterior().
Converted the remaining unguarded ensemble progress output to message().
Updated package documentation for the precipitation-isotope calibration slope.

Initial CRAN release.

Bayesian inversion of leaf-wax δ²H to precipitation δ²H using spatially-aware hierarchical calibrations (14 model variants from Bradley 2026). Posteriors are pre-computed in Stan and shipped as serialized draws, so prediction does not require Stan.
Four-level claim taxonomy via assess_claim(): from analytical-noise thresholds (Level 1) through directional hydroclimate change (Level 2, via corroborating evidence OR the vegetation-only envelope from compute_vegetation_envelope()), quantitative magnitude (Level 3), and unique attribution to precipitation isotopes (Level 4).
Per-record change detection (detect_change()) with autocorrelation-adjusted thresholds and full posterior propagation through invert_d2H(). The within-record detection threshold decomposes the variance of the difference between two single samples as 2 * sigma_residual^2 * (1 - rho_t) + 2 * sigma_analytical^2, applying the lag-1 autocorrelation factor only to the residual term (analytical measurement error is independent between samples).
A 100-draw preview posterior tier ships with the package; complete posterior files are distributed separately.