Developer preview · local traces
Evals Playground
Inspect model usage, latency, failures and cost from the same neutral Nja harness. This first slice is read-only and provider-independent.
Visible runs—
Error rate—
Total tokens—
Estimated cost—
Loading tracesReading local metadata…
| Status | Time | Session | Provider / model | Tokens | Latency | Cost | Note |
|---|
Trace explorer: metadata, tokens, cost, latency, errors and secret redaction.
NextEvaluation sets: opt-in content capture, clone-to-eval, deterministic graders and human annotations.
GateImprovement loop: isolated replay and patch proposal; human approval before apply or merge.