Skip to main content
A capability release object is a harness change plus its evidence, trust state, rollout state, and monitoring history. It moves through a fail-closed lifecycle: gate → promote → monitor → rollback.

Gate

At the gate, your agent reads the freshest Trial Result for the intended treatment, baseline, and task corpus. The gate checks validity, coverage, calibration, contamination, and evidence posture. Missing graders, stale evidence, invalid replay, structural confounds, or an inspect posture block a promotion claim.

Promote

Promotion records that the measured capability may be used for the declared scope. It is not a universal ranking or a promise beyond the corpus and harness that were evaluated. Optimizer finalists must pass a matching holdout before promotion or public claims. A baseline refresh is not promotion: it updates the managed reference evidence without authorizing a new treatment.

Monitor

Monitoring reruns the bounded measurement on a schedule or when the harness, corpus, or operating conditions materially change. Freshness is part of trust: old evidence may remain useful for diagnosis, but it cannot silently stand in for evidence about a changed capability.

Roll back

Rollback withdraws the promoted treatment when monitoring shows regression, degraded evidence, an invalid comparison, or a trust failure. Your agent should state the rollback reason, preserve the receipt, and propose the next bounded repair or comparison rather than overwrite history.

States your agent should name

  • Trust state: whether the evidence is decision-grade, directional, or needs inspection for the declared claim.
  • Rollout state: gated, promoted, monitored, or rolled back.
  • Freshness state: current for the declared inputs, superseded, or stale.
Every transition fails closed. A favorable test headline cannot overcome degraded evidence, an incomplete holdout, a missing grader, or an inspect-only result. Ask your agent to make the state and next action explicit in every report-back.