Autonomy levels

The ladderCopy Anchor Link

AutonomyLevel is an ordered enum. Each level includes the ones below it.

1apiVersion: numinance.dev/v1
2kind: AutonomyPolicy
3metadata: {name: org-ladder, scope: SCOPE_ORG}
4spec:
5  defaults: {level: AUTONOMY_LEVEL_PR}
6  rules:
7✓    - {match: {playbook: dependency-upgrade, risk: RISK_CLASS_LOW}, level: AUTONOMY_LEVEL_MERGE} (matched)
8  floors:
9    rules:
10      - {match: {risk: RISK_CLASS_HIGH}, max_level: AUTONOMY_LEVEL_SUGGEST}
11      - {match: {environment: production}, max_level: AUTONOMY_LEVEL_MERGE}

Defaults, rules and floorsCopy Anchor Link

An AutonomyPolicy has three parts in its spec: defaults, rules and floors. This is the organization example from the spec:

apiVersion: numinance.dev/v1
kind: AutonomyPolicy
metadata: {name: org-ladder, scope: SCOPE_ORG}
spec:
  defaults: {level: AUTONOMY_LEVEL_PR}
  rules:
    - {match: {playbook: dependency-upgrade, risk: RISK_CLASS_LOW}, level: AUTONOMY_LEVEL_MERGE}
  floors:
    rules:
      - {match: {risk: RISK_CLASS_HIGH}, max_level: AUTONOMY_LEVEL_SUGGEST}
      - {match: {environment: production}, max_level: AUTONOMY_LEVEL_MERGE}

• A default is where a change lands if nothing more specific matches.
• A rule matches on playbook, risk and environment and sets a level. Rules can raise or lower the default.
• A floor is a ceiling in practice: it sets a max_level for matching changes. A rule cannot exceed it. In the example, high-risk changes can never go above SUGGEST, and production never above MERGE, whatever a rule says.

Floors are declared with a merge directive that forbids overriding them in lower scopes: a repository policy cannot lift what the organization capped. numi init writes a repository policy at AUTONOMY_LEVEL_PR with no floors block, because floors set below the org scope are refused.

Start supervised, qualify, authorizeCopy Anchor Link

A higher level is not granted because a model got better. The intended path is:

1. Start at a supervised level, typically AUTONOMY_LEVEL_PR, where a human merges every change.
2. Qualify: a playbook can declare qualification thresholds, such as a minimum number of runs and maximum rework, regression and rollback rates, with zero security escapes. These are measured over real runs.
3. Authorize: a human accepts the promotion by changing policy. The agent cannot promote itself, and a failed qualification does not raise a level automatically.

If outcomes degrade, the same path works downward: lower the default or add a floor.

Failure modesCopy Anchor Link

• A rule asks for MERGE on a high-risk change: the floor wins and the change stays at SUGGEST.
• A repository tries to set floors: the document is refused.
• The breaker moves to a more restrictive position: the effective level drops regardless of policy.

Next stepsCopy Anchor Link