What AI evaluation means
AI evaluation is the practice of judging the answers an AI system produces, separately from the system that produced them. A model that writes an answer cannot be trusted to grade it. An evaluation platform keeps the judge apart, fixes the rubric and stores every run so results can be compared over time.
How Latent Lens evaluates an answer
- Instructions: explicit constraints are extracted from the prompt (length, format, JSON, language, things to include or avoid) and checked one by one.
- Claims: the response is split into atomic factual claims.
- Grounding: each claim is matched to your ground truth and labeled supported, partial, contradicted or unsupported. Numeric drift, invented entities and flipped facts are caught deterministically, and an open hallucination model (HHEM) checks meaning.
- Scoring: faithfulness, hallucination rate, instruction adherence, relevance and an overall score are computed in code from those labels, never guessed by a model.
Every report ends with a failure taxonomy and a remediation step per failure type, so your team knows what to change in the prompt, the retrieval or the guardrails.
Why scores computed in code matter
When a model invents the score, the same answer can get 72 today and 81 tomorrow. Latent Lens only lets models assign labels. The maths is fixed and versioned (the rubric version is stamped on every run), so a change in score always means a change in the answer.
Where to start
- Paste one answer into the free Response Checker and see the full report in seconds.
- Read about hallucination detection and RAG evaluation for the details of grounding.
- When one answer is not enough, a Quality Audit runs hundreds of your real conversations.