{ "schema_version": 1, "id": "prosa_pt", "name": "Prosa-PT · rubricas filtradas", "tasks": [ "prosa_pt" ], "expected_counts": [ 1000 ], "family": "prosa", "summary": "Qualidade de respostas a conversas reais em português, com histórico e rubricas binárias.", "score_final_content_only": true, "seed": null, "dataset_rows_in_git": false, "api_judge": true, "protocol": { "language": "pt-BR", "generated_conversations": 1000, "scored_conversations": 981, "frozen_excluded_conversations": 19, "filtered_rubrics": 8408, "selection": "public frozen question_excluded and checklist_filtered_indices; never recomputed", "candidate_history": "original fixed conversation_input; generate only the final assistant response", "candidate_system": "native zero_shot conversation template, unchanged", "candidate_sampling": "explicit model/mode profile, including its output limit; no reasoning_effort", "judge": "gpt-4.1-2025-04-14", "judge_temperature": 0, "judge_max_tokens": 4096, "judge_reasoning_effort": null, "judge_prompt": "native RRD F.3; all original rubrics per included question; final content only", "primary_metric": "micro_filtered_rubric_pass_rate_percent", "secondary_metrics": [ "macro_filtered_question_pass_rate_percent", "empty_final", "truncated" ], "failure_policy": "native: retry failed calls up to three attempts; parse the first returned judgment once; score valid YES/NO selected rubrics and report missing-verdict coverage; require all 1000 generations and judgments for all 981 included conversations", "aggregation_policy": "native_first_judgment_valid_verdicts_v1" }, "category": "chat", "directory": "benchmarks/chat/prosa_pt", "protocol_id": "prosa-pt-native-first-judgment-valid-verdicts-v1", "campaign_enabled": true, "results_output": "results///", "profile_plan": "profiles/plan.json" }