{
  "created_utc": "2026-09-19T06:17:10.573192+00:00",
  "prior_commit": "54215e7",
  "pilot_observations": "Published frozen coefficients reproduce all displayed treatment model coefficients and variance components to rounding. Training-vs-heldout age-14 mean difference 0.7753 days. Heldout reduced-control point estimate 0.6707 versus 0.5823 days full; cannot claim the direction is refuted.",
  "primary_hypothesis": "For the same 33 age-14 training birds, their full-fit age predictions exceed predictions after exclusion of their individual and after exclusion of their known brood of origin by at least 0.5 days on average.",
  "design": "Use all 11,116 deposited candidate CpGs. Main refits retain the authors\u2019 3,184 globally filtered columns so the contrast isolates inclusion in coefficient fitting; controls redo mean/variance filtering inside each training fold. Lambda=0.169141621319677, alpha=0.2 fixed to source values. This is a conditional sensitivity audit, not unbiased selection of a new clock.",
  "solver": "A converged independent elastic-net implementation with gaussian response and feature scaling. Before refits require the original 115-site support and maximum full-training prediction discrepancy <0.03 day versus published coefficients. Full fit was checked: 115 support, max prediction discrepancy 0.0086 day. Coefficient differences in correlated predictors are retained and documented. Frozen external predictions always use exact published coefficients.",
  "primary_exclusions": [
    "Leave-one-individual-out using RNR, including both samples for the sole repeated bird",
    "Leave-one-known-genetic-brood-out for 2018 samples; other samples grouped by RNR because brood IDs were not supplied. This does not imply that unknown broods are independent."
  ],
  "primary_gate": "For BOTH exclusion designs, mean paired (full-fit minus excluded) prediction among the 33 age-14 training birds >=0.5 day; 95% percentile origin-brood bootstrap interval of the paired mean excludes zero; at least 15 known origin-broods represented.",
  "uncertainty": "5,000 origin-brood bootstrap resamples, seed 291919; unit is brood, not CpG. The paired counterfactual uses the same samples.",
  "secondary_outputs": [
    "All 67 out-of-fold predictions and errors, with training age distribution and cohort disclosed",
    "96-sample hybrid treatment estimate: replace only the 33 reused predictions with exclusion-based predictions; external 63 retain published coefficients. This is a sensitivity estimate involving different fitted clocks, not a new independently trained unified clock.",
    "Original full 96, heldout-only 63, known-unseen-brood-only subsets; no equivalence inferred from nonsignificance",
    "Descriptive actual train/heldout errors, variance and all three treatment groups"
  ],
  "controls": [
    "Foldwise mean/variance feature filtering",
    "Lambda multipliers 0.5 and 2.0 with same-individual exclusions; report both, do not choose preferred outcome",
    "Exact published versus converged full predictions to distinguish stopping-rule effects",
    "Original sample-selection status nr_cpg and origin-brood overlap reported",
    "Brood bootstrap for revised treatment contrasts; leave-one-brood influence",
    "Independent verification of one full and one exclusion fit using a second objective solver or native glmnet if feasible"
  ],
  "limits": [
    "Same cohort reused for computational exclusion; no new biological experiment or geographic validation",
    "Chronological age, experiment, year and library composition are partly confounded",
    "No inference that molecular ageing is faster/slower in organismal terms or that the source treatment effect is false",
    "Crossfoster receiving-brood and library metadata may require an additional provenance check; uncertainty cannot ignore available experimental units"
  ],
  "decision": "If primary paired effect fails, record the failure and continue the discovery loop. If it passes, complete controls, second novelty review, note, clean rerun and package before declaring completion."
}
