"""Approved post-stop scoring: all arms share complete canonical components.""" import argparse import json from pathlib import Path import time import numpy as np from gre24_prepare_v2 import sha from r4a_part import contract, exact from s2_custody_common import checked, prediction_inputs from s2_custody_statistics import part_metrics, contrast from s2_recovery_supervisor import save ARMS=('EIK','GRAPH','B0_S2') METRICS=('relative_l2','kendall_tau_b','last_location','near_wall') def summaries(rows, roster, pass_ids): by={r['case']:r for r in rows};groups={} for r in roster:groups.setdefault(r['component_id'],[]).append(r) failed_components={k:dict(parts=[{'case':r['case'],'nodes':r['nodes'],'prediction_pass':r['case'] in pass_ids} for r in members], nodes=sum(r['nodes'] for r in members)) for k,members in groups.items() if any(r['case'] not in pass_ids for r in members)} report={} for metric in METRICS: values={a:{} for a in ARMS};undefined={} for group,members in sorted(groups.items()): if group in failed_components:continue absent=[] for member in members: for arm in ARMS: v=by.get(member['case'],{}).get('metrics',{}).get(arm,{}).get(metric,{}).get('value') if v is None or not np.isfinite(v):absent.append(dict(case=member['case'],arm=arm,metric=metric)) if absent:undefined[group]=absent;continue for arm in ARMS: values[arm][group]=float(np.mean([by[r['case']]['metrics'][arm][metric]['value'] for r in members])) component_ids=sorted(values['EIK']) item=dict(component_ids=component_ids,components=len(component_ids), parts=sum(len(groups[g]) for g in component_ids),nodes=sum(sum(r['nodes'] for r in groups[g]) for g in component_ids), means={a:float(np.mean(list(v.values()))) if v else None for a,v in values.items()}, component_values=values,undefined_components_excluded_from_all_arms=undefined) if metric=='relative_l2': item['contrasts']={} for candidate in ['GRAPH','B0_S2']: delta=contrast(values['EIK'],values[candidate],replicates=2000,seed=2026111408) delta.pop('region',None) delta.update(formula='1 - E_'+candidate+'/E_EIK',reference='EIK',candidate=candidate, interpretation='Positive means candidate has lower error than EIK; descriptive percentile interval, no gate') item['contrasts'][candidate]=delta report[metric]=item return report,failed_components def score(root): c=contract(root);eng=json.loads((root/'continuation-contract.json').read_text()) for p,h in eng['code_sha256'].items():checked(p,h) frozen=root/'pass-predictions-freeze.json';freeze=json.loads(frozen.read_text()) if freeze['contract_sha256']!=sha(root/'contract.json') or freeze['continuation_contract_sha256']!=sha(root/'continuation-contract.json'): raise ValueError('Freeze/contract mismatch') requested={r['case'] for r in c['parts']};passed=set(freeze['parts']);failed={r['case'] for r in freeze['canonical_failures']} if passed&failed or passed|failed!=requested:raise ValueError('Canonical coverage mismatch') for entry in freeze['parts'].values(): checked(entry['result'],entry['result_sha256']);checked(entry['predictions'],entry['predictions_sha256']) # Labels are opened only after every PASS prediction hash has been verified. labels={r['case']:r for r in json.loads(Path(c['label_index']).read_text())} historical={r['case']:r for r in json.loads(Path(c['historical_scores']).read_text()) if r['arm']=='B0'} out=root/'postfit';out.mkdir(exist_ok=True);start=time.process_time();rows=[] for row in c['parts']: case=row['case'] if case not in passed:continue q=prediction_inputs(row);label=labels[case] fold=next(a['fold'] for a in row['assignments'] if a['role']=='test') if label['fold']!=fold:raise ValueError('Cached fold mismatch') with np.load(checked(label['path'],label['sha256']),allow_pickle=False) as z: exact(z['raw_ids'],q['raw_ids'],'target/query IDs');exact(z['counts'],q['counts'],'target/query counts');truth=z['seconds'].copy() if not np.isfinite(truth).all() or np.any(truth<0):raise ValueError('Invalid cached truth') entry=freeze['parts'][case] with np.load(checked(entry['predictions'],entry['predictions_sha256']),allow_pickle=False) as z: exact(z['raw_ids'],q['raw_ids'],'prediction IDs');exact(z['counts'],q['counts'],'prediction multiplicity') metrics={arm:part_metrics(z[array],truth,q['counts'],q['near_wall'],q['raw_ids'],q['coordinates_over_Rg'],int(case.split('_')[0])) for arm,array in [('EIK','F'),('GRAPH','G')]} b0=historical.get(case) if b0 is None:raise ValueError('Missing historical B0 row') metrics['B0_S2']={metric:b0[metric] for metric in METRICS} rows.append(dict(case=case,component_id=row['component_id'],fold=fold,metrics=metrics, cached_label_sha256=label['sha256'],prediction_sha256=entry['predictions_sha256'])) save(out/'progress.json',dict(stage='CACHE_ONLY_SCORING',parts=len(rows),worker_cpu_seconds=time.process_time()-start)) report,excluded=summaries(rows,c['parts'],passed) save(out/'part-scores.json',dict(rows=rows,predictions_freeze_sha256=sha(frozen))) save(out/'score-summary.json',dict(status='SCORED_DESCRIPTIVE_COMPLETE_COMPONENTS',authorizing_ruling='1794830693.515879', requested_parts=len(requested),requested_components=len({r['component_id'] for r in c['parts']}), pass_prediction_parts=len(passed),canonical_failures=freeze['canonical_failures'],excluded_components=excluded, summary=report,delta_direction='EIK reference; 1 - GRAPH/EIK and 1 - B0_S2/EIK', superseded_draft_direction='1 - EIK/GRAPH was not scored; changed by Claude1794830693', historical_B0_interpretation='whole distance construction; not pure propagation attribution', new_decodes=0,new_fits=0,worker_cpu_seconds=time.process_time()-start, predictions_freeze_sha256=sha(frozen),part_scores_sha256=sha(out/'part-scores.json'), limitations=['Fixed v5 geometry and S2 query domain, not fresh blind testing', 'Failed components excluded identically from every arm; no trimmed membership', 'Component-only descriptive bootstrap; no CI or 10% gate', 'Historical B0 differs in full distance construction, not just propagation'])) if __name__=='__main__': p=argparse.ArgumentParser();p.add_argument('root',type=Path);a=p.parse_args();score(a.root)