Reference-free evaluation lets you score agent quality when you do not have labelled answers.
Beyond exact-match
We combine rubric-based LLM judges with behavioural checks so you can track quality drift even on open-ended tasks where no labelled answer exists. Reference-free evaluation scores what an agent actually did, not how closely it matched a golden string.
Signals that work without labels
- Faithfulness: does the answer stay grounded in the retrieved context?
- Task completion: did the agent reach the goal state its tools imply?
- Self-consistency: do repeated runs converge, or scatter?
- Rubric judging: a calibrated judge model scores tone, safety, and completeness.
Putting it into practice
Start by sampling real production traffic, define a short rubric per task type, and let Nexora run the judges on every trace automatically. You get a trend line for quality you can alert on — long before a user files a complaint. Pair it with a small human-labelled set to keep the judges honest.