The metrics
| Metric | What it measures |
|---|---|
| Faithfulness | Share of claims backed by the ground truth: (supported + half of partial) / claims |
| Hallucination rate | Share of claims that contradict the source or have no support in it |
| Instruction adherence | Explicit instructions from the prompt that were followed, out of those checked |
| Relevance | How much of the question the answer covers |
| Overall | A weighted mix; any high-severity failure caps it at 74 |
Model-agnostic by design
Latent Lens judges the output, not the model. It works the same on answers from GPT, Claude, Gemini, Llama, Mistral or your own fine-tune, so you can compare them on equal terms.
Reproducible by design
- The judge only labels; scores are computed in code.
- The rubric and engine versions are stamped on every run.
- The same input always gives the same score, so a change in score means a change in the answer.
Compare runs, not impressions
Every check you run while signed in is saved to your dashboard. Pick two runs to see score changes and which failures were fixed or are new. Export everything as CSV or JSON for your own analysis.
Read more about hallucination detection or start with the free checker.