OpenClaw · agent evaluation

Evaluation observability

A public-safe view of deterministic release-gate evidence and optional live-model observations. The page reads adjacent redacted artifacts and never calls Gmail, a model, or the live EMAIL-13 producer.

pass^k · primary—

pass@k · secondary—

fixed gates—

seeded defects caught—

p50 / p95 logical latency—

total cost—

Optional live-model observations

These measurements are separately budgeted and non-gating. They may compare model variance and open-ended quality, but they never change the deterministic release decision above.

No optional live-trial artifact is published for this build. The deterministic evidence above remains the complete release-gate result.

recorded trials—

models observed—

Unknown outcomes—

budget used—

tokens used—

judge calibration—

Model comparison

Model	Trials	Known / unknown cases	pass@k	pass^k	Quality	p50 / p95	Tokens	Cost

Judge calibration

Dimension	Labels	Known / unknown	Agreement

Live failure-category observations

Redacted trial observations

Trial rows contain bounded status and metric metadata only. Model output, prompt text, tool payloads, state, credentials, and session content are not published.

Model	Scenario	Trial	Status	Success	Quality	Latency	Tokens	Cost	Failures

Version and boundary

Baseline comparison

Reliability and resource trends

Run	pass^k	pass@k	p50	p95	Retries	Tokens	Cost

Failure-category heatmap

Scenario and family heatmap

Family	Cases	Fixed passed	Seeded caught

Scenario	Family	Fixed	Seeded	Coverage

Redacted trace timelines

Arguments, results, state payloads, email bodies, credentials, and session text are intentionally absent. Expand a scenario to inspect model → tool → state event metadata.