EH
Eval Healthcheckbeta
Pre-release health check

Check whether your AI eval signal is trustworthy before release.

Upload an eval dataset, run health checks on your judge and dataset, and get a release-readiness report your team can act on.

Built for AI engineers running evals before release.

Sample report verdict
WARN82 / 100 · Mostly Stable
Review needed before using this eval as a release gate.
6 of 30 samples produced unstable verdicts across 5 runs and the dataset has 3 rows with missing expected values.
What gets checked
  • Verdict stability across N runsLLM-assisted
  • Required dataset columnsDeterministic
  • Missing or empty expected valuesDeterministic
  • Duplicate rowsDeterministic
  • Basic length outliersDeterministic

Start with judge and dataset health

3 live
⚗
AI-assisted
01Judge Reliability
Is your LLM judge consistent enough to trust? We re-run it multiple times per sample and surface unstable verdicts.
🗄
Deterministic
02Dataset Health
Is your eval dataset structurally usable and obviously risky? Deterministic checks for missing fields, duplicates, and shape.
📈
Deterministic
03Score Health
Is score movement meaningful or likely noise? Compare eval score changes against judge reliability.