Results
5 configs on the dev split (100 questions). Unfiltered, every column is the harness's summary. Filters recompute retrieval metrics in the browser from per-question scores; the other columns stay whole-split summaries.
100 of 100 questions
Leaderboard
mean [95% CI], percentage points| H1L Lemmatised BM25 + dense bge-m3 over structure chunks, fused by RRF (local twin of H1). | 55.9 [46.1–64.7] | 41.1 [33.4–49.2] | 38.5 [31.8–45.2] | 17 ms | 19 ms |
|---|---|---|---|---|---|
| D1 Naive RAG - dense bge-m3 over fixed 512-word windows. | 52.4 [43.3–60.7] | 37.4 [30.0–44.9] | 36.2 [29.4–42.4] | 16 ms | 18 ms |
| T1L H1L over table-row chunks and XLSX cell rows (local twin of T1, no reranker). | 50.6 [41.2–58.9] | 38.5 [30.8–46.1] | 35.2 [28.4–42.0] | 17 ms | 21 ms |
| B1 BM25 over raw tokens, fixed 512-word windows. | 42.0 [32.7–50.6] | 37.6 [29.4–46.4] | 31.6 [24.3–38.6] | 0.1 ms | 0.1 ms |
| B2 BM25 over MorphoDiTa lemmas with diacritics folded, fixed 512-word windows. | 39.6 [30.2–47.7] | 34.5 [26.7–42.6] | 30.0 [22.6–37.2] | 0.3 ms | 1 ms |
Retrieval only: accuracy, citation precision, conflict, abstention and cost columns appear once an answer layer has run.
Config × category
| Config | Lookup | Numeric | Cross lingual | Multi hop | Conflict | Decoy | Cim tie out | Version | Unanswerable | Clause set |
|---|---|---|---|---|---|---|---|---|---|---|
| B1· BM25 over raw tokens, fixed 512-word windows. | 85 | 50 | 0 | 33 | 50 | 25 | 48 | 40 | · | 25 |
| B2· BM25 over MorphoDiTa lemmas with diacritics folded, fixed 512-word windows. | 69 | 50 | 0 | 33 | 40 | 33 | 52 | 40 | · | 25 |
| D1· Naive RAG - dense bge-m3 over fixed 512-word windows. | 69 | 61 | 58 | 50 | 42 | 25 | 44 | 60 | · | 34 |
| H1L· Lemmatised BM25 + dense bge-m3 over structure chunks, fused by RRF (local twin of H1). | 92 | 72 | 33 | 47 | 50 | 44 | 48 | 40 | · | 36 |
| T1L· H1L over table-row chunks and XLSX cell rows (local twin of T1, no reranker). | 92 | 61 | 33 | 35 | 37 | 44 | 48 | 40 | · | 36 |
Config × messiness
a question carries the tags of its evidence documents · cells recomputed from per-question gold ranks| Config | cz en mix | cz number format | draft version | eur tab | hidden sheet | jpeg artifacts | merged header | near duplicate | quoted reply | scan | signed value | skew | split table | unit kc | unit mil | unit tis |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| B1· BM25 over raw tokens, fixed 512-word windows. | 53 | 39 | 50 | 33 | 33 | 39 | 36 | 46 | 38 | 39 | 40 | 39 | 23 | 33 | 49 | 40 |
| B2· BM25 over MorphoDiTa lemmas with diacritics folded, fixed 512-word windows. | 50 | 36 | 50 | 33 | 33 | 36 | 36 | 44 | 38 | 36 | 38 | 36 | 23 | 33 | 48 | 37 |
| D1· Naive RAG - dense bge-m3 over fixed 512-word windows. | 60 | 50 | 56 | 50 | 50 | 51 | 55 | 56 | 55 | 51 | 48 | 51 | 38 | 50 | 54 | 50 |
| H1L· Lemmatised BM25 + dense bge-m3 over structure chunks, fused by RRF (local twin of H1). | 65 | 53 | 55 | 48 | 48 | 52 | 53 | 53 | 55 | 52 | 51 | 52 | 38 | 48 | 53 | 53 |
| T1L· H1L over table-row chunks and XLSX cell rows (local twin of T1, no reranker). | 58 | 49 | 52 | 45 | 45 | 47 | 44 | 52 | 38 | 47 | 47 | 47 | 31 | 45 | 51 | 49 |