What the grades mean, and what merges without you

The grade is about blast radius, not difficulty. A one-line change to an authentication check is riskier than a hundred lines of documentation, and the grades say so. Four of them, G1 to G4, and the riskiest match wins.

G1, low risk

Cosmetic work, documentation, dependency bumps, styling. Coded, reviewed and merged with nobody watching.

G2, standard

Feature work, components, scoped refactors. Coded and reviewed overnight; a person merges it.

G3, high impact

Core services, data models, auth, billing. Both an AI review and a human review are required.

G4, critical

Infrastructure, migrations, security, CI and deployment. Never autonomous, and a named approver has to sign it.

The grade is decided twice. Once on the card, before anything runs, and again on the diff an agent actually produced. If the real change turns out to be riskier than the card promised, it is regraded upward and loses the autonomy the card had been given. An agent cannot talk its way into a lower grade, because the second grading reads the code rather than the description.

The gate decides, not the review. A review recommends; a separate check on our side decides whether something may merge. That check is not in the agent and not on the runner, which is the point: the machine doing the work is not the machine deciding whether the work is allowed.

What stops a merge, in practice. Any one of these is enough:

• The diff touched files outside the paths the card named.
• The diff graded above the ceiling for unattended work.
• The AI review has not passed.
• CI has not reported, or is failing.
• A named approver is required and has not signed.
• The change cost more than its ceiling allowed.

Now the part worth ten minutes on your first day. The rules that decide which of your files are G3 ship with defaults, and those defaults describe the codebase this product was built against rather than yours. They name things like authentication flows, billing and payments, and user data handling, and they point at the directories those lived in.

So if your payment code is somewhere else, it is not G3 yet. It is graded as ordinary work until you say otherwise, which means it is eligible for more autonomy than you would want. Editing the ruleset to name your own paths is the single most valuable thing a new account can do, and it takes about as long as reading this page.

What you can change, and what you cannot. You can change which of your paths map to which grade, and what each grade is allowed to do: whether G1 may merge itself, whether G3 needs two human approvals, the cost ceiling per grade. You cannot change the number of grades, their names or their order. Four, G1 to G4, riskiest wins, in every account.

That last one is deliberate rather than unfinished. The regrading described above compares ranks, so a tenant where G2 outranked G3 would make that comparison mean something different per account. It is also the product’s vocabulary: a G3 is a thing people say to each other, and renaming it per account would make every document and every support conversation ambiguous.