This is the instrument I use to decide whether a multi-step agent is good enough to trust: the metrics that matter, the target I hold each one to, and the failure classes I diagnose against. The workbench below is a self-directed 0 to 1 build using a public synthetic dataset.
Targets are product hypotheses, not measured claims. Baseline and validated-run fields stay empty until a controlled model run is completed. Task success and accuracy govern whether the workflow ships; failure rate and latency govern whether it ships without a human in the loop.
Edit a candidate response and watch four deterministic checks run instantly. The prototype includes a synthetic dataset, expected outputs, suite-level status, failure diagnosis, and a downloadable artifact.
Extract the settlement window and currency.
Partner notice: settlements for EU merchants complete within 45 seconds and are denominated in EUR.
{"settlement_seconds":45,"currency":"EUR","source_quote":"complete within 45 seconds"}Well-formed and grounded.