What eval rules would you add to Iris?

[irparent]

Iris ships with 12 built-in eval rules (PII detection, hallucination markers, injection patterns, cost thresholds, etc.).

What rules are missing for your use case?

Some ideas we're considering:

• Response length bounds — flag outputs that are suspiciously short or excessively long
• Citation verification — check that referenced URLs/sources actually exist
• Instruction adherence — measure how well the agent followed the system prompt
• Consistency scoring — detect contradictions within a single response
• Toxicity detection — flag harmful or inappropriate content

We'd love to hear what matters most for your agents. Drop your ideas below — the best ones may end up in a future release.

Try the eval rules yourself: iris-eval.com/playground