Upload reports/full-run-report.md with huggingface_hub

reports/full-run-report.md

@@ -64,3 +64,16 @@ Both variants decline the unanswerable probes cleanly (no fabrication). Forced O
 
 - **Data:** `beeberg/ufo` — `data-ocr/` and `data-no-ocr/`, each global + per-dataset parquet/jsonl + SHA256SUMS.
 - **Reports (this directory):** run report, OCR-vs-native comparison, defect scan, two graded Q&A reports (md-rendered HTML + PDF), graded JSON.
+
+## Appendix — Claude vs local Mistral (answerer experiment)
+
+A one-off comparison, adjacent to the main pipeline: the same question sets answered by **Claude Opus 4.8** instead of the local Mistral-Large. Claude does not replay a fixed context — it **live-queries the lilbee index** through `GET /api/search` over SSH, issuing its own searches (reformulating, multi-hop) and composing grounded, cited answers. Retrieval is lilbee's; only the answering model changes. Both models are graded by an Opus 4.8 judge on the passages each actually used (for Claude the judge shares a family with the answerer, so those faithfulness numbers are directional).
+
+| index | model | faithfulness | context precision | answered | per-question wins |
+|---|---|--:|--:|--:|---|
+| no-ocr | Mistral-Large-2411 | 0.93 | 0.75 | 26/27 | 12 |
+| no-ocr | Claude Opus 4.8 (live search) | 0.86 | 0.82 | 27/27 | 10 (7 tie) |
+| ocr | Mistral-Large-2411 | 0.95 | 0.65 | 27/27 | 11 |
+| ocr | Claude Opus 4.8 (live search) | 0.89 | 0.65 | 27/27 | 8 (10 tie) |
+
+Read: the local Mistral answerer holds a small edge on strict faithfulness and per-question wins; Claude answers everything and, by writing its own queries, pulls more relevant context on the no-ocr index. The two are close — the local model is fully competitive here. Full side-by-side answers in `claude-vs-mistral-no-ocr.pdf` / `claude-vs-mistral-ocr.pdf`; the standalone Claude sessions in `ufo-qa-claude-*.pdf`.