collect-generation.py

1"""Read-only extraction: stdout is a public metadata record; no verse text/logs."""2import argparse3import collections4import hashlib5import json6from pathlib import Path78parser = argparse.ArgumentParser()9parser.add_argument('--source-output', type=Path, required=True)10args = parser.parse_args()11root = Path(__file__).resolve().parents[3]12audit_path = root / 'cli/translate-bible/reports/nichijo-v1-numbering-alignment.json'13raw_audit = audit_path.read_bytes()14audit = json.loads(raw_audit)15models = collections.Counter()16dates = []17matched = missing = changed = records = undated = 018for name, expected in audit['input_files'].items():19    path = args.source_output / name20    if not path.is_file():21        missing += 122        continue23    raw = path.read_bytes()24    if hashlib.sha256(raw).hexdigest() != expected:25        changed += 126        continue27    matched += 128    for verse in json.loads(raw)['verses'].values():29        records += 130        models[verse.get('model', 'unrecorded')] += 131        if verse.get('timestamp'):32            dates.append(verse['timestamp'])33        else:34            undated += 135usage = audit['usage']36stages = [37    ('initial_alignment', usage['initial_alignment']['batches']),38    ('repair_proposals', usage['repair_round']['batches']),39    ('repair_verification', usage['repair_round']['independent_verification']['usage']),40]41fields = ['input_tokens', 'cache_creation_input_tokens', 'cache_read_input_tokens', 'output_tokens']42rows = []43for stage, batches in stages:44    entries = []45    for batch in batches:46        entry = {'record': batch.get('batch', batch.get('file')),47                 'response_sha256': batch.get('response_sha256', batch.get('sha256')),48                 'model_labels': [batch['model']] if 'model' in batch else list(batch['modelUsage']),49                 'tokens': {field: batch['usage'][field] for field in fields},50                 'is_error': batch.get('is_error'),51                 'terminal_reason': batch.get('terminal_reason')}52        entries.append(entry)53    totals = {field: sum(e['tokens'][field] for e in entries) for field in fields}54    rows.append({'stage': stage, 'requests': len(entries), 'tokens': totals,55                 'total_tokens': sum(totals.values()), 'records': entries})56print(json.dumps({57    'captured_at': '2026-10-03',58    'audit_sha256': hashlib.sha256(raw_audit).hexdigest(),59    'generation': {'scope': 'Only local source chapter files whose SHA-256 matches the release audit input_files.',60        'matched_chapter_files': matched, 'missing_chapter_files': missing, 'changed_chapter_files': changed,61        'source_records': records, 'models': dict(models), 'earliest_record_timestamp': min(dates),62        'latest_record_timestamp': max(dates), 'records_without_timestamp': undated,63        'timezone': 'Not recorded in timestamp strings; do not infer UTC.',64        'total_generation_tokens': None,65        'limitations': 'Saved verse timestamps, not a complete run start/end log. Source records are not final aligned release verses. No whole-project model or token claim.'},66    'review_usage': {'stages': rows, 'total_tokens': sum(r['total_tokens'] for r in rows),67        'scope': 'Recorded alignment, repair-proposal and repair-verification requests only; includes unsuccessful recorded requests. Excludes original generation, Codex work and unrecorded calls.',68        'counting': 'Sum top-level input_tokens + cache_creation_input_tokens + cache_read_input_tokens + output_tokens once per recorded request. Do not add nested iterations, modelUsage or thinking tokens again.',69        'dates': None, 'model_label_note': 'Labels as recorded in the audit, not independently verified vendor model identities.'},70    'website': {'created_on': '2026-10-03', 'tool': 'Codex', 'model': None, 'tokens': None,71        'note': 'Exact underlying model and session token total are not supplied in the site evidence.'}72}, ensure_ascii=False, indent=2))73