Changelog

referent 0.1.0

Initial scoped release of the CDF-first normative-modelling API.

• Gaussian and SHASH distributional reference fits, total coefficient uncertainty, response transforms, support checks, site adaptation, PIT recalibration, harmonisation, joint scores, and FDR-aware flags.
• Fail-closed model selection now requires valid fits and complete marginal, tail, shape, and conditional calibration; uncalibrated exploratory selection requires an explicit opt-in.
• Training, calibration, and evaluation roles carry full SHA-256 provenance. Calibration is bound to its uncertainty estimand and coefficient-draw count.
• Longitudinal dynamics, transitions, and forecasts share one recorded marginal-uncertainty estimand and disclose plug-in kernel uncertainty.
• Frozen references use bundle schema 1.0.0 with validated ref_write() and ref_read() migration for unversioned development bundles. Because RDS is executable serialization, ref_read() requires an explicit trusted = TRUE acknowledgement before deserializing a bundle from an authenticated source.
• Numerical validation includes offline PCNtoolkit 1.3.0 oracles, a replicated five-scenario release matrix, HBR/site receipts, NHANES cohort diagnostics, cross-platform R CMD check, and scale/allocation budgets.

The 0.1.0 release-candidate evidence passes its registered computational gates without relaxing them. Four replicated PCNtoolkit scenarios are equivalent and the fresh-seed skew-heavy matrix is superior with a valid comparator; deterministic high-acceptance HBR sampling passes convergence and leave-one-site-out-selected adaptation passes every site. The NHANES 2017-2018 rerun passes computationally but is explicitly post-hoc model-development evidence because its earlier failure informed the scale model. A separately preregistered, untouched NHANES 2013-2014 confirmation passes the unchanged full calibration contract and is the evidence used for the scoped conditional-transport claim.