Evaluation laboratory
Real generated measurements. Explicit providers. Bounded claims.
Measure the investigator. Keep the claims bounded.
Reproducible labeled synthetic incidents measure failure-domain identification, evidence retrieval, safe abstention, and tool integrity.
Generated benchmark
Benchmark not yet generated.
This page reads artifacts/eval/latest.json. No scores appear until the evaluation harness has produced a real artifact.
Deterministic provider results evaluate this pipeline on authored synthetic cases. They do not establish live model accuracy, customer usage, or production MTTR reduction. Latency is synthetic benchmark investigation latency; no human-operator study has been performed. AI-Assisted Investigator comparisons appear only after an actual live-provider evaluation. The legacy invalid-reference metric counts nonexistent evidence IDs only; it does not measure whether a cited record supports a claim.