benchmark_report.md

# Synthetic RAG Benchmark Report

- Generated at: `2026-03-06T15:45:57.457782+00:00`
- Seed: `1337`
- Corpus docs: `1500`
- Queries: `300`
- Auto-tune: `True`
- Retrieval config: `{'step_k_ratio': 0.06, 'min_step_k': 20, 'min_match_ratio': 0.7, 'final_match_ratio': 0.65, 'noise_weight': 0.01, 'ensemble_top_k': 2, 'ensemble_vote_threshold': 0.4}`

## Aggregate Metrics

| Method | Avg Precision | Avg Recall | Avg F1 | Exact Match Rate |
|---|---:|---:|---:|---:|
| naive | 0.5161 | 0.6567 | 0.4254 | 0.1167 |
| instructed | 1.0000 | 0.3849 | 0.4627 | 0.2333 |
| verified_structural | 1.0000 | 0.3849 | 0.4627 | 0.2333 |

| verified_structural_ensemble | 1.0000 | 0.4590 | 0.5268 | 0.3333 |

| verified_consensus | 1.0000 | 0.4372 | 0.5026 | 0.3167 |

## Notes

- `naive`: single partition `[n]`.
- `instructed`: heuristic partition `[n-1, 1]` for `n > 1`.
- `verified_structural`: non-oracle structural scorer (balance + step penalty + chunking heuristics).
- `verified_structural_ensemble`: non-oracle vote across top structural partitions.
- `verified_consensus`: non-oracle self-consistency scorer (retrieval stability + structure + match density).

No theorem discharge is claimed by this benchmark.