nx_conformal.nx

nx_conformal.nx -- distribution-free prediction intervals. Vovk/Gammerman/Shafer 2005 ("Algorithmic Learning in a Random World"): for any black-box predictor f, conformal prediction builds a prediction SET around f(x_new) with EMPIRICAL COVERAGE guarantee. No distributional assumption. No model retraining. Given calibration set {(x_i, y_i)} of size n and a non- conformity score s_i = |y_i - f(x_i)|: q_alpha = ceil((n+1)(1-alpha)) / n quantile of {s_i} Predicted set for x_new: [f(x_new) - q_alpha, f(x_new) + q_alpha] Guarantee (Vovk Theorem 2.1): P(y_new in predicted_set) >= 1 - alpha for exchangeable data. The structural answer to "i want to predicitably make an 18 year old human whether male or female or have you not hallucinate into c, i cant get that in the current setup." Without conformal, model says "age 22" but true age is some random thing. WITH conformal: "age 22; with 90% coverage, true age in [19, 27]" (distribution-free, model-agnostic). Two modes: SPLIT -- standard split-conformal (Papadopoulos 2002): calibration set held out; gives a single q_alpha. ADAPTIVE -- locally-weighted conformal (Lei/Wasserman 2014): nonconformity scores weighted by feature density; narrower intervals where data is dense. (Stub -- returns SPLIT result with weights=1 for v1.) genealogy_id: vovk_2005_papadopoulos_2002 lineage_id: substrate_conformal_v1 5W+H+GLP linkage: per-prediction emit VERDICT record with what=VERDICT, how=conformal, performance=interval_width_q10.