Autonomy engine

How Curia’s autonomy score governs skill execution through action_risk levels, audit trails, and planned automatic adjustment.

This page covers the governance mechanics of Curia’s autonomy engine — how the score controls what skills can execute, how changes are audited, and what’s planned for automatic adjustment. For the conceptual overview and calibration guidance, see Autonomy.

How action_risk connects to autonomy

Every skill declares an action_risk level in its manifest. This field establishes the minimum autonomy score at which that skill can be invoked without explicit confirmation.

action_risk value	Minimum score	Skill examples
none	0	Web search, reading email, summarizing content
low	60	Writing to memory, updating contacts
medium	70	Sending email, sending Signal messages
high	80	Creating calendar events, making commitments on your behalf
critical	90	Financial actions, bulk deletions, irreversible operations

At the current Phase 1 implementation, the autonomy score governs Curia’s behavior through prompt injection — the behavioral description for the active band is included in the system prompt on every task. Phase 2 will add hard execution gates that block skill invocations when the live score is below the skill’s declared floor.

The action_risk field is required on all skill manifests. Curia will not start if any skill omits it.

Score history

Every change to the autonomy score is written to an append-only history record in Postgres. No entry is ever updated or deleted. The history records:

• The new score and band
• The previous score
• Who made the change (ceo or, in a future release, system for automatic adjustments)
• An optional reason you provided
• The timestamp

This gives you a complete audit trail of how Curia’s independence has evolved over time, and why.

Automatic score adjustment

The autonomy scoring pass runs as part of the dream engine and evaluates Curia’s recent actions to adjust the autonomy score based on performance metrics:

• Competence (45%) — task success rate, accuracy of information retrieved, follow-through on commitments
• Commitment (35%) — reliability on recurring tasks, consistency with stated preferences
• Compatibility (20%) — alignment with your communication style and decision patterns

Automatic adjustments require a minimum of 30 actions before any change, apply time-decay weighting so recent behavior matters more, and cannot increase the score if factual errors or overconfidence penalties are active. All automatic changes are written to the audit record with the reason noted, so you can see exactly what changed and why.

Skills and action_risk

See how skill risk levels are declared in the manifest.

Audit & logging

The append-only audit trail that records all system actions.