# DataTrace leaderboard

Ranked by **grounded accuracy** (RC@1^g): the root cause is correct and evidence
recall is at least 0.5. `infl` is the **evidence gap** (RC@1 − RC@1^g), or the
share of all runs that are correct but have insufficient evidence. `LG` reports
the same unsupported-correct rate. Brackets show 95% bootstrap confidence
intervals.

| # | model | n | RC@1^g [95% CI] | RC@1 | infl | LG [95% CI] | Fix | ER | ER|c | EF1 | cost | steps |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| 1 | `moonshotai-kimi-k3` | 11 | **0.818** [0.55,1.00] | 1.000 | 0.182 | 0.182 [0.00,0.46] | 0.727 | 0.757 | 0.757 | 0.500 | 596 | 20.1 |
| 2 | `openai-gpt-5.6-sol` | 128 | **0.586** [0.50,0.67] | 0.688 | 0.102 | 0.102 [0.06,0.16] | 0.320 | 0.816 | 0.806 | 0.611 | 2858 | 31.4 |
| 3 | `anthropic-claude-sonnet-5` | 128 | **0.523** [0.44,0.61] | 0.633 | 0.109 | 0.109 [0.06,0.16] | 0.305 | 0.686 | 0.633 | 0.564 | 1470 | 19.6 |
| 4 | `z-ai-glm-5.2` | 52 | **0.481** [0.35,0.61] | 0.712 | 0.231 | 0.231 [0.12,0.35] | 0.385 | 0.762 | 0.732 | 0.532 | 7978 | 23.6 |
| 5 | `deepseek-deepseek-v4-flash` | 128 | **0.344** [0.27,0.43] | 0.414 | 0.070 | 0.070 [0.03,0.12] | 0.398 | 0.841 | 0.750 | 0.514 | 7959 | 43.9 |
| 6 | `x-ai-grok-4.5` | 128 | **0.234** [0.16,0.31] | 0.305 | 0.070 | 0.070 [0.03,0.12] | 0.141 | 0.821 | 0.737 | 0.506 | 4527 | 38.3 |
| 7 | `google-gemini-3.5-flash` | 128 | **0.164** [0.10,0.23] | 0.195 | 0.031 | 0.031 [0.01,0.06] | 0.266 | 0.840 | 0.761 | 0.550 | 3123 | 23.0 |

_Grounded accuracy is the ranking metric because RC@1 alone does not distinguish
between supported and insufficiently supported diagnoses. A larger `infl` means
that more runs found the correct root cause without reaching the evidence-recall
threshold._