"""Frozen full-domain RMSE/max(truth) summaries on identical paired support.""" import numpy as np def summarize(rows, jobs): per = [] for j in jobs: chosen = [r for r in rows if r['split'] == j['id']] if {r['case'] for r in chosen} != set(j['test']) or len(chosen) != len(j['test']): raise ValueError('Complete paired prediction support required') valid = [r for r in chosen if r['A'] is not None and r['H2'] is not None] summaries = {} for method in ['A', 'H2']: values = np.asarray([r[method] for r in valid], np.float64) if not np.isfinite(values).all() or np.any(values < 0): raise ValueError('Invalid paired metric') summaries[method] = {name: float(fn(values)) if len(values) else None for name, fn in [('min',np.min),('median',np.median),('max',np.max)]} delta = np.asarray([r['H2']-r['A'] for r in valid],np.float64) per.append(dict(split=j['id'],requested_test_parts=315,paired_predictions=len(chosen), scored_pairs=len(valid),missing_original=list(j['missing_test']), undefined=[r['case'] for r in chosen if r['A'] is None or r['H2'] is None], summaries=summaries,paired_mean_H2_minus_A=float(delta.mean()) if len(delta) else None, paired_median_H2_minus_A=float(np.median(delta)) if len(delta) else None)) if len(per) != 5 or {r['split'] for r in per} != set(range(5)): raise ValueError('Exactly five original splits required') means = {method: {name: float(np.mean([r['summaries'][method][name] for r in per])) if all(r['summaries'][method][name] is not None for r in per) else None for name in ['min','median','max']} for method in ['A','H2']} return dict(per_split=per,mean_of_five_split_summaries=means, direction='Negative paired H2-minus-A favours H2; descriptive only', descriptive_paired_uncertainty=paired_uncertainty(rows),significance_gate=False,metric='RMSE / max(full truth)', full_requested_315_coverage=all(not r['missing_original'] for r in per)) def paired_uncertainty(rows): grouped = {} visits = {} for r in rows: visits[r['case']] = visits.get(r['case'], 0) + 1 if r['A'] is not None and r['H2'] is not None: grouped.setdefault(r['case'], []).append(r['H2']-r['A']) cases = sorted(grouped) values = np.asarray([np.mean(grouped[x]) for x in cases], np.float64) if not np.isfinite(values).all(): raise ValueError('Nonfinite part-level paired difference') counts = dict(candidate_unique_parts=len(visits), candidate_part_visits=dict(sorted(visits.items())), undefined_parts=sorted(set(visits)-set(cases))) if not len(values): return dict(defined=False, unique_parts=0, point=None, ci95=None, gate=False, **counts) rng = np.random.default_rng(2026111408) draws = values[rng.integers(0,len(values),size=(2000,len(values)))].mean(axis=1) return dict(defined=True, unique_parts=len(values), point=float(values.mean()), **counts, ci95=np.quantile(draws,[.025,.975],method='linear').tolist(), replicates=2000, seed=2026111408, interval='percentile_linear', unit='unique_part_mean_of_split_visit_differences', part_visits={x:len(grouped[x]) for x in cases}, gate=False, limitation='Descriptive part bootstrap; random-part splits may share geometry families')