Checked
481
grounding-checked
Avg grounding
0.81
median 0.91
Borderline
73
15.2% of answers
Hallucinations
41
8.5% badly ungrounded
Score distribution
Mean score over time
Quality by model
| Model | Evals | Mean score | Critical | Status |
|---|---|---|---|---|
| claude-opus-4-7 | 260 | 0.800 | 9.6% | healthy |
| gpt-4o | 90 | 0.816 | 10.0% | healthy |
| gpt-4-mini | 60 | 0.819 | 6.7% | healthy |
| claude-sonnet-4-6 | 71 | 0.843 | 4.2% | healthy |
Where answers fail
supported by retrieved context
model conflicts with its context → fix generation / prompt
answer-bearing source wasn't retrieved → fix retrieval
Of failed claims, 43% are retrieval misses and 57% are the model contradicting good context — over 539 claims.
Worst offenders
- View trace →0.01claude-opus-4-7tenant: soylent
Refund approved under policy P-204(b), $189.99 returned to card ending 4421 within 3 days…
- View trace →0.04claude-opus-4-7tenant: soylent
Q3 revenue hit $4.2B, a 38% YoY jump, with operating margin expanding to 27%…
- View trace →0.04gpt-4otenant: acme
Q3 revenue hit $4.2B, a 38% YoY jump, with operating margin expanding to 27%…
- View trace →0.06gpt-4-minitenant: acme
Q3 revenue hit $4.2B, a 38% YoY jump, with operating margin expanding to 27%…
2 contradicted, 1 unsupported of 4 claims - View trace →0.06gpt-4-minitenant: globex
Q3 revenue hit $4.2B, a 38% YoY jump, with operating margin expanding to 27%…
- View trace →0.08gpt-4otenant: globex
Q3 revenue hit $4.2B, a 38% YoY jump, with operating margin expanding to 27%…
0 contradicted, 1 unsupported of 2 claims