"""D3 oracle diagnostic, not a calibration; only frozen prediction/label caches.""" import json from pathlib import Path import time import numpy as np from freeze_g1_components import sha from s2_recovery_supervisor import save from s2_custody_common import read_json,checked,test_rows,expected_models,prediction_path from s2_custody_statistics import component_values ROOT=Path('artifacts/s2-posthoc-diagnostics-2026-11-15') EVAL=Path('artifacts/s2-custody-evaluation-2026-11-15') def decompose(prediction,truth,counts): p=np.asarray(prediction,dtype=np.float64);t=np.asarray(truth,dtype=np.float64);c=np.asarray(counts,dtype=np.float64) if p.ndim!=1 or p.shape!=t.shape or p.shape!=c.shape or not len(p):return dict(valid=False,reason='empty or invalid shape') if not all(np.isfinite(v).all() for v in [p,t,c]) or np.any(c<=0) or np.any(c!=np.floor(c)): return dict(valid=False,reason='nonfinite eligible item or invalid counts') ps=float(np.max(np.abs(p)));ts=float(np.max(np.abs(t))) if ts==0:return dict(valid=False,reason='zero_truth',e2=None,H=None,S=None,a=None) if ps==0:return dict(valid=False,reason='zero_pred',e2=1.,H=None,S=None,a=None) # Unit-vector projection avoids squaring tiny P or subtracting close energies. wp=np.sqrt(c)*(p/ps);wt=np.sqrt(c)*(t/ts) pn=np.linalg.norm(wp);tn=np.linalg.norm(wt);u=wp/pn;v=wt/tn cosine=float(u@v) beta_ld=np.longdouble(ps)/np.longdouble(ts)*np.longdouble(pn)/np.longdouble(tn) beta=float(beta_ld) a_ld=np.longdouble(cosine)/beta_ld a=float(a_ld) if abs(a_ld)<=np.finfo(float).max else None e2=float(np.sum((beta*u-v)**2));H=float(np.sum((cosine*u-v)**2));S=float((beta-cosine)**2) if not np.isfinite([e2,H,S]).all():return dict(valid=False,reason='unrepresentable normalized energy',a=a) if not np.isclose(e2,H+S,rtol=1e-10,atol=1e-12):raise ArithmeticError('Orthogonal decomposition failed') return dict(valid=True,reason=None,a=a,a_scientific=np.format_float_scientific(a_ld,precision=17), e2=e2,H=H,S=S,identity_absolute_error=abs(e2-H-S), sign_reversal=cosine<0,nonpositive_a=cosine<=0,zero_cross=cosine==0, perfect_prediction=e2==0,negative_prediction_items=int(np.sum(p<0))) def synthetic_checks(): for p,t,want in [([2,4],[1,2],(.5,0.,1.,1.)),([1,2],[1,2],(1.,0.,0.,0.)), ([-1,-2],[1,2],(-1.,0.,4.,4.)),([1,0],[0,1],(0.,1.,1.,2.)), ([1e-200,2e-200],[1,2],(1e200,0.,1.,1.))]: d=decompose(p,t,[1,1]);assert d['valid'] np.testing.assert_allclose([d['a'],d['H'],d['S'],d['e2']],want,rtol=1e-12,atol=1e-12) z=decompose([0,0],[1,2],[1,1]);assert not z['valid'] and z['reason']=='zero_pred' and z['e2']==1 and z['H'] is None for p,t,c in [([1,2],[0,0],[1,1]),([],[],[]),([1,np.nan],[1,2],[1,1])]:assert not decompose(p,t,c)['valid'] p=np.array([.1,.4,.9]);t=np.array([.2,.3,1.]);c=np.array([3,1,7]);d=decompose(p,t,c) np.testing.assert_allclose(d['a'],np.sum(c*p*t)/np.sum(c*p*p),rtol=1e-12) np.testing.assert_allclose(d['e2'],np.sum(c*(p-t)**2)/np.sum(c*t*t),rtol=1e-12) return 10 def main(): start=time.process_time();assert not (ROOT/'D3-started.json').exists(),'No duplicate diagnostic pass' tests=synthetic_checks();c=read_json(EVAL/'contract.json');original=read_json(EVAL/'custody/result.json') for name in ['label-index.json','part-scores.json']:checked(EVAL/'custody'/name,original['output_sha256'][name]) predindex=read_json(checked(EVAL/'prediction-index.json',read_json(EVAL/'predictions-complete.json')['prediction_index_sha256'])) labels={r['case']:r for r in read_json(EVAL/'custody/label-index.json')} original_scores={(r['case'],r['arm'],r['pool'],r['seed']):r for r in read_json(EVAL/'custody/part-scores.json')} save(ROOT/'D3-started.json',dict(rulings=['1794731518.379529','1794731544.355059','1794731547.266819'], description='oracle diagnostic, not a calibration',synthetic_checks=tests,code_sha256=sha(__file__), zero_pred='e2=1 retained as original error; a/H/S undefined and excluded from valid decomposition', tolerance=dict(rtol=1e-10,atol=1e-12),new_labels_decoded=False)) rows=[];metric_rows=[];maximum_identity_error=maximum_frozen_error=0. for fold in range(5): for row in test_rows(c,fold): label=labels[row['case']];checked(label['path'],label['sha256']) with np.load(label['path'],allow_pickle=False) as z:truth=z['seconds'];ids=z['raw_ids'];counts=z['counts'] for model in expected_models(c,fold): p=prediction_path(EVAL,model,row['case']);checked(p,predindex[str(p)]) with np.load(p,allow_pickle=False) as z: assert np.array_equal(z['raw_ids'],ids) and np.array_equal(z['counts'],counts) d=decompose(z['seconds'],truth,counts) old=original_scores[row['case'],model['arm'],model['pool'],model['seed']]['relative_l2']['value'] if d['valid']: error=abs(d['e2']-old**2);maximum_frozen_error=max(maximum_frozen_error,error) assert np.isclose(d['e2'],old**2,rtol=1e-10,atol=1e-12) maximum_identity_error=max(maximum_identity_error,d['identity_absolute_error']) record=dict(case=row['case'],fold=fold,component_id=row['component_id'],arm=model['arm'],pool=model['pool'],seed=model['seed'],**d) rows.append(record) metric_rows.append(dict(case=row['case'],arm=model['arm'],pool=model['pool'],seed=model['seed'], **{key:dict(value=d[key] if d['valid'] else None,reason=d['reason']) for key in ['e2','H','S']}, nonpositive=dict(value=float(d['nonpositive_a']) if d['valid'] else None,reason=d['reason']), reversal=dict(value=float(d['sign_reversal']) if d['valid'] else None,reason=d['reason']))) roster=read_json(c['roster'])['parts'];members={p for f in read_json(checked(c['exposure_membership'],c['exposure_membership_sha256'])) for p in f['parts']} tables=[];comparisons=[] for subset in ['all','not_known_exposed']: for pool in [25,50,100]: byarm={} for arm in ['N','E','R2','B0']: values={};details={};excluded={} for metric in ['e2','H','S','nonpositive','reversal']: values[metric],excluded,details=component_values(metric_rows,roster,arm,pool,metric,subset!='all',members) keys=set.intersection(*(set(v) for v in values.values()));means={k:float(np.mean([v[x] for x in sorted(keys)])) if keys else None for k,v in values.items()} if keys:assert np.isclose(means['e2'],means['H']+means['S'],rtol=1e-10,atol=1e-12) tables.append(dict(subset=subset,pool=pool,arm=arm,components=len(keys),parts=sum(details[k]['parts'] for k in keys), excluded_components=excluded,mean=means,oracle_removable_squared_error_share=means['S']/means['e2'] if keys and means['e2']>0 else None, warning='oracle diagnostic, not a calibration; mean squared error is not square of primary mean L2')) byarm[arm]=(values,details) for ref,cand in [('N','E'),('R2','N'),('B0','N')]: aa,details=byarm[ref];bb,_=byarm[cand];keys=set(aa['e2'])&set(bb['e2']) am={k:float(np.mean([aa[k][x] for x in sorted(keys)])) for k in ['e2','H','S']} bm={k:float(np.mean([bb[k][x] for x in sorted(keys)])) for k in ['e2','H','S']} comparisons.append(dict(subset=subset,pool=pool,reference=ref,candidate=cand,components=len(keys), parts=sum(details[x]['parts'] for x in keys),reference_means=am,candidate_means=bm, candidate_minus_reference={k:bm[k]-am[k] for k in am},warning='squared-error diagnostic difference, not primary Delta')) save(ROOT/'D3-part-decomposition.json',rows);save(ROOT/'D3-tables.json',tables);save(ROOT/'D3-comparisons.json',comparisons) summary=dict(status='D3_COMPLETE',rows=len(rows),valid=sum(r['valid'] for r in rows),undefined=[r for r in rows if not r['valid']], nonpositive_a_rows=sum(r.get('nonpositive_a',False) for r in rows),sign_reversal_rows=sum(r.get('sign_reversal',False) for r in rows), zero_cross_rows=sum(r.get('zero_cross',False) for r in rows),maximum_identity_absolute_error=maximum_identity_error, maximum_frozen_squared_error_difference=maximum_frozen_error,cpu_seconds=time.process_time()-start, new_labels_decoded=False,predictions_modified=False,calibration=False,synthetic_checks=tests, output_sha256={p.name:sha(p) for p in ROOT.glob('D3-*.json')}) save(ROOT/'D3-receipt.json',summary) print(json.dumps(summary));print(json.dumps([x for x in tables if x['subset']=='all' and x['pool']==100])) if __name__=='__main__':main()