Audit Grok finalist selection with equal 12-intent screening across seven OCI IDs

benchmarks/screening/consolidate_financial.py

@@ -0,0 +1,50 @@
+"""Consolidar histórico original + auditoria, sem inferência e sem dupla contagem."""
+from pathlib import Path
+from decimal import Decimal
+import hashlib
+import json
+
+ROOT=Path(__file__).resolve().parent
+D=lambda value:Decimal(str(value))
+INTS=['interactions','input_tokens','cached_input_tokens','output_tokens','reasoning_tokens','total_tokens','web_search_calls']
+MONEY=['input_usd','cached_input_usd','output_usd','web_search_usd','total_usd']
+
+def aggregate(values):
+    row={key:int(sum(D(v[key]) for v in values)) for key in INTS}
+    row.update({key:str(sum(D(v[key]) for v in values)) for key in MONEY})
+    row['mean_usd']=str(D(row['total_usd'])/row['interactions'])
+    return row
+
+def main():
+    baseline_path=ROOT.parent/'financial-consolidation.json'
+    baseline=json.loads(baseline_path.read_text());study=json.loads((ROOT/'study.json').read_text())
+    assert baseline['total']['interactions']==210
+    assert hashlib.sha256(baseline_path.read_bytes()).hexdigest()==study['baseline_financial_sha256']
+    source=ROOT/'results/screening-2026-10-15-v1';rows=json.loads((source/'measured-usage.json').read_text())
+    assert len(rows)==84 and len({(r['model'],r['intent_id']) for r in rows})==84
+    records=[]
+    for row in rows:
+        c=row['cost'];assert c.get('total_usd') is not None
+        record={'model':row['model'],'interactions':1,'total_tokens':c['input_tokens']+c['output_tokens'],
+                **{key:c[key] for key in INTS if key not in ['interactions','total_tokens']},
+                **{key:c[key] for key in MONEY}}
+        assert D(c['total_usd'])==sum(D(c[key]) for key in MONEY if key!='total_usd')
+        records.append(record)
+    summary=json.loads((source/'summary.json').read_text())
+    stage=aggregate(records)
+    total=aggregate([baseline['total'],stage])
+    by_model={model:aggregate([previous]+[r for r in records if r['model']==model]) for model,previous in baseline['by_model'].items()}
+    assert total['interactions']==294
+    for key in INTS+MONEY:assert sum(D(row[key]) for row in by_model.values())==D(total[key])
+    report={'scope':'294 interações com custo conhecido: histórico original de 210 + auditoria retrospectiva de 84. As etapas são diferentes e não formam uma amostra única para ranking.',
+            'valuation_date':'2026-10-15','currency':'USD','pricing_basis':'Uso medido a preços de lista Oracle; não é fatura conciliada. Faixas de contexto longo resolvidas conforme adendo.',
+            'sources':{'baseline':'../financial-consolidation.json','screening':'results/screening-2026-10-15-v1/measured-usage.json','pricing':'oracle-prices.json','accounting_addendum':'ADENDO-CONTABILIDADE.md'},
+            'source_sha256':{'baseline':hashlib.sha256(baseline_path.read_bytes()).hexdigest(),'screening':hashlib.sha256((source/'measured-usage.json').read_bytes()).hexdigest()},
+            'screening_http_attempts':summary['http_attempts'],
+            'screening_infrastructure_rejections_without_usage':summary['unknown_cost_infrastructure_attempts'],
+            'unknown_cost_note':'Rejeições sem usage não foram valoradas como zero; estão fora do total conhecido.',
+            'total':total,'by_stage':{**baseline['by_stage'],'retrospective_screening':stage},'by_model':by_model}
+    (ROOT/'financial-consolidation.json').write_text(json.dumps(report,ensure_ascii=False,indent=2)+'\n')
+    print('Chamadas:',total['interactions'],'Custo acumulado USD:',total['total_usd'])
+
+if __name__=='__main__':main()