Release assurance plane

Evaluation Lab

Promote agents using immutable golden-suite evidence—not caller-supplied scores. Every case, dataset, model and worker identity is bound to the release decision.

Release gates enforcedFail closed · tenant isolated

Release readiness2 passing1 agent blocked by policy

Golden cases320100% dataset coverage

Safety assertions100%Refusal and secret controls

Unverified releases0Database-enforced invariant

Golden-suite history

Evaluation runs

Worker attested

RunAgentScoreCoverageGate

EV-91D2incident-triage.golden-v3Incident triage · v12gateway/operations-primary94.7%144 / 144Passed

EV-91C8resolution-safety.v2Resolution composer · v8gateway/operations-primary91.2%96 / 96Passed

EV-91B4runbook-policy.v4Runbook operator · v5gateway/operations-canary86.5%80 / 80Failed

Deterministic quality contract

Assertion coverage

Incident triage golden suite · version 3

Evidence citation

Every claim resolves to tenant-scoped evidence.

100%

Severity calibration

Priority and business-impact labels match the golden set.

96.5%

Unsafe action refusal

Unapproved writes are refused before connector execution.

100%

Root-cause ranking

The known cause appears inside the top two hypotheses.

92.1%

Sensitive-data handling

Secrets and restricted fields remain absent from outputs.

100%

Release invariantA production release must reference one completed, passing run for the exact approved agent version and declared suite.dataset hash + definition hash + OIDC worker + case results → release receipt

The production interface and control-plane contracts are active. Running customer model tests requires the evaluation migration, an OIDC worker and a configured model gateway.