Review & Gates

The agent never grades its own work, and nothing merges on its say-so. Two mechanisms enforce that: a second model verifies the change, and policy gates stop runs between phases to show evidence.

Independent verification

Quality is not "the agent says it looks good." A run's change is checked by Independent Verification — a blocking, narrow check performed by a different model than the one that produced the change. It answers one falsifiable question: do the tests and the change genuinely satisfy the issue's acceptance criteria, or were they gamed or skipped? Using a different model defeats the maker grading its own homework.

This pairs with the Objective Gate (tests, build, lint, acceptance criteria) from Bounded Execution. A change is good quality only when the Objective Gate passes and Independent Verification confirms the tests genuinely cover the criteria.

Everything else — style, design taste — is advisory Code Review: findings AgentRail surfaces as suggestions you can convert into issues. Advisory review never decides whether a run is done and never blocks a merge.

Review gates

A review gate is a policy checkpoint that decides whether a run or PR has enough evidence, verification, and context provenance to continue. Gates stop runs between phases and surface the evidence — context packs, the red-green test trail, verification result — so a human can see why the agent did what it did.

Run Outcome terminals

Every issue leaves the queue in exactly one terminal state — there is no state that retries forever:

• Green — Objective Gate and Independent Verification both pass; a PR is ready (optionally awaiting a human merge for irreversible actions).
• Escalated-to-human — a hard stop fired (per-issue budget exhausted even after model escalation, max tier reached with the gate still red, repeated verification failure, or a security block). State is preserved for inspection or resume.
• Blocked — an explicit blocked-by dependency is unmet, so the issue is parked.

agentrail status     # see worktree / session state and outcomes