Design an autopilot with a constitution
Let AI propose and agents act without surrendering goals, evidence, permissions, or manual control.
The operating model
Yapture separates interpretation, decision, permission, and execution:
| Layer | Responsibility |
|---|---|
| Optional AI | Propose attributes, evidence mappings, and explanations |
| User policy | Define goals, weights, phases, hard constraints, and autonomy |
| MAUT kernel | Reconcile evidence and rank eligible alternatives deterministically |
| Capability client | Execute only authorized actions after policy gates pass |
This is how autopilot remains defined by the user rather than by an opaque “clanker” ranking.
Autonomy modes
Start with the least powerful mode:
Off → Explain → Suggest → Reorder → Schedule → Act
- Explain scores the existing order without changing it.
- Suggest shows a proposed ordering while manual order remains primary.
- Reorder applies a ranked view with immediate undo.
- Schedule proposes calendar blocks for confirmation.
- Act lets a scoped agent perform an approved action.
Never escalate a mode because the user ignored a suggestion or used the product frequently.
Policy shape
interface AutonomyPolicy {
mode: 'off' | 'explain' | 'suggest' | 'reorder' | 'schedule' | 'act';
allowedActions: string[];
maxActionsPerRun?: number;
requireConfirmationAbove: 'none' | 'low' | 'medium' | 'high';
acceptedLearningSignals: Array<'explicit-pairwise' | 'confirmed-reorder' | 'rating'>;
}
The decision policy should also record its scope, goal, attributes, weights, time phases, and hard constraints.
Execution gate
An action may proceed only when every condition passes:
eligible by hard constraints
AND allowed by user autonomy policy
AND allowed by capability or OAuth scope
AND feature is actually available
AND confirmation threshold is satisfied
AND rate/action limit is not exceeded
MAUT determines fit among eligible alternatives. It does not grant permission.
Audit record
Record enough information to reproduce and challenge a decision:
- explicit reference time,
- case and policy versions/hashes,
- effective weights and phase,
- ranked alternatives and confidence,
- excluded tasks and reasons,
- proposed and confirmed actions,
- model/SDK version,
- evidence references without copying unnecessary sensitive content.
An audit trail is still user data. Give it retention and deletion behavior rather than keeping it forever by default.
Safe preference learning
Accepted signals:
- “I prefer A to B.”
- A confirmed manual reorder.
- An explicit rating.
Rejected passive signals:
- Hover duration.
- An abandoned drag.
- Merely viewing a task.
- Agent behavior not confirmed by a person.
Provide Reset learned preferences independently of task-history deletion.
Anonymous lists
For an accountless list, keep decision preferences on the device by default. An owner capability is not a stable personal identity, and two capability holders should not have their preference models merged silently.
If the list is later claimed, ask whether to import, keep local, or discard the device-local model.
MAGYC.fit
MAGYC.fit grading support is Coming soon. Until a versioned grading contract exists, Yapture may export a preview evidence package but must not display a fake grade or invent a submission endpoint.
A MAUT score is relative utility under a policy. It is not a grade of a person or their productivity.