{
  "created_utc": "2026-09-19T06:26:54.992036+00:00",
  "frozen_protocol_commit": "16c5e39",
  "timing": "After failed Python full-fit support calibration and successful native R full-fit reproduction; before any exclusion predictions were produced or inspected.",
  "reason": "Converged sklearn selects 115 sites but four differ from the published set; maximum training prediction difference 0.008582 day. Native glmnet 5.0 defaults reproduce all published coefficients within 1.1e-14, including the exact support. The frozen exact-support gate therefore failed for Python.",
  "amendment": "Use native glmnet 5.0 default solver as primary for all prespecified exclusions and sensitivities. Keep every primary effect-size, direction, bootstrap and sample-count gate unchanged. Retain tight-convergence native fits and independent Python KKT-certified fits as numerical controls, explicitly separate from exclusion effects.",
  "supplemental_metadata": "Dryad 261370 supplies origin, rearing, cross-foster Pair and sequencing Library for all 96 test birds, including all 33 age14 training birds. Target has 22 origin broods, 22 rearing broods, 15 pairs and 8 libraries. Treatment labels agree after mapping original Other to later Control; no Enlarged/Reduced discrepancies.",
  "added_controls": [
    "Leave entire cross-foster Pair out of training (other experiments use RNR); also leave rearing brood out.",
    "Report pair-clustered as well as origin-brood-clustered 5000-resample intervals for paired age14 differences, seed291919.",
    "Treatment sensitivity: origin-brood REML matching source plus Pair-cluster bootstrap for OLS treatment contrasts, with and without fixed library effects. Report all versions; no preferred significance selection.",
    "Leave one cross-foster Pair out influence for original, heldout and hybrid treatment contrasts.",
    "Describe CpG coverage (nr_cpg), library and treatment composition by training membership."
  ],
  "limits": "Supplemental analyses do not change the primary gate. Pair bootstrap quantifies dependence of measured exclusion shifts conditional on the fitted corpus, not uncertainty over new training populations. Existing hyperparameter and candidate filtering reuse remain explicit."
}
