Stage 6 Audit
生成后审计页。它把 5-agent review 中的硬约束转成可检查计数；任何非预期计数都应在进入综述生成前处理。
| check | value | expected/action |
|---|---|---|
| paper_master_map_rows | 29 | expected_29_for_panorama_generation |
| read_priority_rows | 29 | must cover every kept paper |
| verified_metric_rows | 0 | 0 means no SOTA/ranking conclusion allowed |
| verified_schema_matches_template | true | paper_metric_result_verified.csv uses verified template fields |
| review_queue_rows | 84 | review_only/do_not_rank |
| method_parse_suspect_rows | 282 | must not enter verified ranking |
| dataset_negated_rows | 1 | negated train/use evidence must not become train facts |
| dataset_bibliographic_rows | 79 | references/related work default to reference_only |
| sota_eligible_missing_group | 0 | must be 0 |
| review_rows_without_source_row_id | 0 | must be 0 |
| current_claim_to_baseline_metric_rows | 136 | must not support claims; baseline context only |
| public_metric_evidence_redacted | true | 6.4 public evidence snippets redact numeric tokens |
Hard Rules
paper_metric_result_verified.csv为空时，6.4/6.5/sota_rankings 不得展示可见 value preview，也不得产出 SOTA/best/outperforms 系统结论。method_parse_suspect=yes的 metric row 只能用于人工查表，不能支撑 claim。evidence_polarity=negated的 dataset evidence 不能升为 train/use 事实。- 公开 markdown 的 review preview 必须隐藏 value；真实数值只保留在 CSV。
claim_metric_match_status=target_mismatch的行只能作为 baseline/comparison context，不能支撑 current-method claim。