"""Preregistered array-only statistics; no dataset or model access.""" from collections import defaultdict import numpy as np from s2_models.objectives import relative_l2 def six_region(low,high,margin=.1): if not np.isfinite([low,high,margin]).all() or low>high or margin<=0:raise ValueError('Invalid interval') if high<0:return 'harmful' if low<=0:return 'no meaningful gain' if high=i lo=np.minimum(i,j);hi=np.maximum(i,j) st=np.sign(t[hi]-t[lo]);sp=np.sign(p[hi]-p[lo]) denominator=np.sqrt(float(np.count_nonzero(st))*float(np.count_nonzero(sp))) if denominator==0:return dict(value=None,reason='zero tau-b denominator') return dict(value=float(np.sum(st*sp)/denominator),reason=None) def last_location(truth,prediction,raw_ids,coordinates_over_Rg): t=np.asarray(truth);p=np.asarray(prediction);ids=np.asarray(raw_ids) a=np.flatnonzero(t==t.max());b=np.flatnonzero(p==p.max()) ia=a[np.argmin(ids[a])];ib=b[np.argmin(ids[b])] return float(np.linalg.norm(np.asarray(coordinates_over_Rg)[ia]-np.asarray(coordinates_over_Rg)[ib])) def part_metrics(prediction,truth,counts,near_wall,raw_ids,coordinates_over_Rg,prefix): p=np.asarray(prediction);t=np.asarray(truth) if p.shape!=t.shape or p.ndim!=1 or not len(p) or not np.isfinite(p).all() or not np.isfinite(t).all():raise ValueError('Invalid scores') near=np.asarray(near_wall,dtype=bool) return dict(relative_l2=relative_l2(p,t,counts,np.ones(len(t),bool)), near_wall=relative_l2(p,t,counts,near) if near.any() else dict(value=None,reason='empty near-wall domain'), kendall_tau_b=kendall_pairs(t,p,prefix), last_location=dict(value=last_location(t,p,raw_ids,coordinates_over_Rg),reason=None)) def component_values(rows,roster,arm,pool,metric='relative_l2',not_known_exposed=False,frozen_members=None): """Complete fixed seed blocks; exposure filters parts before component means.""" if any(type(r['exposure']['development_exposed']) is not bool for r in roster): raise ValueError('development_exposed must be bool; UNKNOWN is a separate inventory field') if not_known_exposed and frozen_members is None:raise ValueError('Frozen exposure membership required') main_excluded=component_values(rows,roster,arm,pool,metric,False)[1] if not_known_exposed else {} parts=[r for r in roster if r['included'] and (not not_known_exposed or r['case'] in frozen_members)] grouped=defaultdict(list) for r in parts:grouped[r['component_id']].append(r) by_key={(r['case'],r['arm'],r['pool'],r['seed']):r for r in rows} values={};excluded={};details={} seeds=[0,1,2] if arm in ['N','E'] else [None] model_pool=None if arm=='B0' else pool for component,members in sorted(grouped.items()): block=[];reasons=[] for member in members: per_seed=[] for seed in seeds: row=by_key.get((member['case'],arm,model_pool,seed)) value=None if row is None else row[metric]['value'] if value is None:reasons.append(dict(case=member['case'],seed=seed,reason='missing prediction/model' if row is None else row[metric]['reason'])) else:per_seed.append(value) block.append(per_seed) details[component]=dict(parts=len(members),nodes=sum(x['nodes'] for x in members), fold=next(a['fold'] for a in members[0]['assignments'] if a['role']=='test')) if component in main_excluded:excluded[component]=dict(**details[component],reasons=main_excluded[component]['reasons'],main_analysis_invalid=True) elif reasons:excluded[component]=dict(**details[component],reasons=reasons) else:values[component]=float(np.mean(block)) return values,excluded,details def contrast(reference,candidate,replicates=2000,seed=2026111408): ids=sorted(set(reference)&set(candidate)) if not ids:return dict(value=None,reason='no complete paired components',components=0) a=np.array([reference[k] for k in ids]);b=np.array([candidate[k] for k in ids]) if a.mean()<=0:return dict(value=None,reason='zero reference mean',components=len(ids)) rng=np.random.Generator(np.random.PCG64(np.random.SeedSequence(seed)));sample=rng.integers(0,len(ids),size=(replicates,len(ids))) aa=a[sample].mean(axis=1);bb=b[sample].mean(axis=1) undefined=np.flatnonzero(aa<=0) if len(undefined):return dict(value=float(1-b.mean()/a.mean()),reason='undefined bootstrap replicates; no replacement or interval', components=len(ids),undefined_replicates=len(undefined),undefined_replicate_indices=undefined.tolist(),low=None,high=None,region=None) boot=1-bb/aa;low,high=np.percentile(boot,[2.5,97.5],method='linear') return dict(value=float(1-b.mean()/a.mean()),low=float(low),high=float(high),components=len(ids),component_ids=ids, reference_mean=float(a.mean()),candidate_mean=float(b.mean()),replicates=replicates,seed=seed,undefined_replicates=0, interval='component-only percentile95 conditional on fixed folds/models/seeds',region=six_region(low,high))