"""Read-only Track A reuse/coverage entry: JSON metadata and opaque byte hashes only. No numpy, model loading, target-array access, fitting, prediction or scoring. """ import ast import hashlib import json import time from pathlib import Path BASE = Path('/home/frank14f/optane/MeltMind') OLD = BASE/'artifacts/gre24-v4-reproduction-2026-11-16/protocol' OUT = BASE/'artifacts/track-a-existing-input-audit-2026-11-18' def read(p): return json.loads(Path(p).read_text()) def sha(p): h=hashlib.sha256() with Path(p).open('rb') as f: for block in iter(lambda:f.read(2**20),b''): h.update(block) return h.hexdigest() def main(): cpu=time.process_time();wall=time.monotonic() if OUT.exists(): raise FileExistsError('Existing audit: inspect, do not overwrite') split_path=OLD/'splits.json';splits=read(split_path);tc=read(OLD/'training-contract.json') prep=read(OLD/'preparation-freeze.json');models=read(OLD/'models-freeze.json');pred=read(OLD/'predictions-freeze.json') hpath=BASE/'artifacts/phase2-h2-pilot-2026-11-17/contract.json';h2=read(hpath) assert sha(split_path)==tc['splits_sha256']=='534b2a25f92dbac2335b6282e1dfe0bb8ecdbc8db55bdfc24c7cffa83f9e931b' assert sha(OLD/'preparation-freeze.json')==tc['preparation_freeze_sha256'] assert sha(OLD/'training-contract.json')==models['training_contract_sha256'] assert sha(OLD/'models-freeze.json')==pred['models_freeze_sha256'] for name,digest in tc['code_sha256'].items(): assert sha(BASE/name)==digest,name assert len(splits)==5 and len(prep['successful'])==615 and len(prep['failed'])==14 cols=next(ast.literal_eval(n.value) for n in ast.parse((BASE/'tools/gre24_features_v2.py').read_text()).body if isinstance(n,ast.Assign) and any(isinstance(t,ast.Name) and t.id=='COLUMNS' for t in n.targets)) assert len(cols)==19 for case,rec in prep['successful'].items(): metadata=Path(rec['output'])/'result.json' assert sha(metadata)==rec['result_sha256'],case assert read(metadata)['features']['columns']==list(cols),case success=set(prep['successful']);hc=set(h2['parts']);universe=set(splits[0]['train'])|set(splits[0]['test']) assert len(universe)==629 and len(hc)==552 and hc<=universe and hc<=success shared=[];mismatches=[] for case in sorted(hc): old=prep['successful'][case];new=h2['parts'][case] names=['features.npy','raw_ids.npy','mass.npy','Vf.npy','seconds.npy','target.npy'] differences=[n for n in names if old['output_sha256'][n]!=new['output_sha256'][n]] if old['nodes']!=new['nodes']:differences.append('nodes') if old['t_scale_seconds']!=new['t_set']:differences.append('t_set') if differences:mismatches.append(dict(case=case,fields=differences)) shared.append(dict(case=case,differences=differences,metadata_hash_comparison_only=True)) rows=[];hashes={};prediction_count=0 for sp in splits: i=sp['index'];assert len(sp['train'])==314 and len(sp['test'])==315 and not set(sp['train'])&set(sp['test']) assert set(sp['train'])|set(sp['test'])==universe entry=models['fits'][str(i)+'/GBRT'];assert entry['status']=='PASS' for key in ['checkpoint','guard_receipt']: p=Path(entry[key]);hashes[str(p)]=sha(p);assert hashes[str(p)]==entry[key+'_sha256'] guard=read(entry['guard_receipt']);assert guard['status']=='PASS' and not guard['monitor_errors'] fitdir=OLD/'fits'/f'split-{i}'/'GBRT';fit=read(fitdir/'result.json');log=read(fitdir/'training-target-access.json') assert read(fitdir/'xgboost-config.json')['version']==[2,1,4] params=fit['parameters'];assert params['n_estimators']==100 and params['max_depth']==8 and params['learning_rate']==.1 assert params['random_state']==sp['seed'] expected=[case for case in sp['train'] if case in success] assert [x['case'] for x in log]==expected assert fit['train_geometries']==len(expected) for item in log: rec=prep['successful'][item['case']] assert item['feature_sha256']==rec['output_sha256']['features.npy'] assert item['target_sha256']==rec['output_sha256']['target.npy'] prows=[r for r in pred['rows'] if r['split']==i] assert len(prows)==315 and {r['case'] for r in prows}==set(sp['test']) for row in prows: if row['case'] not in success:assert row['status']!='PREDICTED';continue assert row['status']=='PREDICTED' value=row['predictions']['GBRT'];p=Path(value['path']);hashes[str(p)]=sha(p);assert hashes[str(p)]==value['sha256'] rec=prep['successful'][row['case']] assert value['nodes']==rec['nodes'] and row['raw_ids_sha256']==rec['output_sha256']['raw_ids.npy'] prediction_count+=1 row=dict(split=i,seed=sp['seed'],roles={}) for role in ['train','test']: cases=set(sp[role]);row['roles'][role]=dict(requested=len(cases),A_valid=len(cases&success),A_missing=sorted(cases-success),H2_case_id_available=len(cases&hc),H2_missing=sorted(cases-hc),both_available_by_metadata=len(cases&hc&success),H2_missing_but_A_valid=sorted((cases&success)-hc),valid_A_cases=sorted(cases&success),existing_H2_cases=sorted(cases&hc)) rows.append(row) assert [r['roles']['test']['A_valid'] for r in rows]==[304,311,308,309,305] assert prediction_count==1537 core=['contract.json','training-contract.json','splits.json','preparation-freeze.json','models-freeze.json','predictions-freeze.json','score-summary.json'] audit=dict(status='A_REUSE_ENTRY_CHECK_PASS',approval='1794973608.321679',protocol_ruling='1794973277.152329',old_source_pins_match=True,columns=list(cols),normalization='Original Eq12 geometry scalings only; no learned standardization; features cast FP32 for XGBoost; target seconds/t_set, predictions converted back to seconds; no t_set input column',splits=rows,old_requested=629,old_success=615,old_failures=prep['failed'],H2_case_id_available=552,H2_missing_total=77,H2_missing_A_success=sorted(success-hc),H2_v4_v5_metadata_mismatches=mismatches,shared_case_checks=shared,shared_metadata_identity_only=True,models_hash_verified=5,A_prediction_files_hash_verified=prediction_count,opaque_file_hashes=hashes,metadata_hashes={str(OLD/f):sha(OLD/f) for f in core}|{str(hpath):sha(hpath)},A_refit_needed_under_frozen_protocol=False,old_summary_context_only=read(OLD/'score-summary.json')['summary']['GBRT']['rrmse_maximum'],no_new_scoring=True,no_target_arrays_opened=True,no_model_deserialization=True,no_fit_or_prediction=True,CPU_seconds=time.process_time()-cpu,wall_seconds=time.monotonic()-wall) OUT.mkdir();(OUT/'audit.json').write_text(json.dumps(audit,indent=2)+'\n') print(json.dumps({k:v for k,v in audit.items() if k in ['status','H2_case_id_available','H2_missing_total','H2_v4_v5_metadata_mismatches','models_hash_verified','A_prediction_files_hash_verified','CPU_seconds','wall_seconds']})) if __name__=='__main__':main()