# Scribe Grounding Leaderboard

Source-Presence Index (SPI) over the checksum-locked grounded suite (24 tasks). Higher is better: SPI 1.00 means every factual sentence carries a valid `[n]` citation (or the model correctly refuses when the sources don't cover the question). No LLM judge is involved — the metric is deterministic.

_Generated 2026-07-17T11:23:56 · suite `b8dbabb216fc`_

| Rank | Model | SPI | Invalid citations |
| ---: | :--- | ---: | ---: |
| 1 | qwen3-32b (groq) | 0.885 | 0 |
| 2 | gemma-4-12b (local) | 0.865 | 0 |
| 3 | gemma-4-E4B (po-stari) | 0.865 | 0 |
| 4 | gpt-oss-120b (groq) | 0.839 | 0 |
| 5 | llama-3.3-70b (groq) | 0.760 | 0 |
| 6 | gemma-4-E2B (po-hub) | 0.673 | 0 |
| 7 | llama-3.1-8b (groq) | 0.451 | 0 |