PAA Progressive Autonomy Architecture earned / scoped / revocable
v0.2.0-draft

Progressive Autonomy Architecture

Let systems earn autonomy.

PAA is an architecture for changing how much authority an automated system has over a bounded task, using recorded evidence and reversible decisions.

Autonomy is earned from observed results, scoped to the task those results describe, and revocable when performance degrades.

The governing idea

Evidence changes oversight

The PAA core governance loop Bound task: Task Bound task Evaluate work: Evaluator Evaluate work Accumulate evidence: Evidence Accumulate evidence Decide: Decision Decide Change oversight: Oversight Change oversight
  • Governance loop
  • Regression
Evidence can make a task eligible for less oversight. Changing its authority is a separate decision. A confirmed regression can restore oversight directly.
  1. A worker performs a bounded task.
  2. An external evaluator judges the result.
  3. The system retains those judgments as evidence.
  4. A governance decision changes oversight when the evidence justifies it.
  5. Failure restores the safer operating mode.

Evidence supports a decision. It does not grant authority by itself.

The task is the unit of governance. A system may earn less review for routine refunds while remaining fully gated for unusual refunds or account changes. Evaluation continues after oversight changes, and degraded results restore the declared fallback.

One concrete task

Refund approval

Every refund is reviewed at first. The review results accumulate across cases. After 200 qualifying decisions clear the declared threshold, an operator may approve automatic handling for routine refunds. Ambiguous cases still escalate, and one confirmed policy failure returns the task to mandatory review.

Refund reference build