Progressive autonomy is not novel. The pattern already operates across several automation
domains. PAA defines a vendor-neutral implementation architecture that is not tied to any
one of them.
SAE J3016 establishes the familiar lineage of discrete automation levels defined by the
allocation of responsibility between a person and an automated driving system. Its levels
are specific to driving and to an operational design domain.
PAA applies the level-based frame to defined tasks and records changes in oversight.
Hellert, Montenegro, and Sulc describe plan-first agentic control at the Advanced Light
Source. The system produces a human-readable execution plan before touching hardware;
human review and policy checks gate execution.
Osprey demonstrates a typed boundary with blocking authorization. It does
not define a general progression and regression lifecycle for reducing that oversight.
Zhang and colleagues describe advancement from human-on-the-loop operation to bounded
autonomy after statistically sufficient incident-free operation. An incident or constraint
violation causes regression, followed by investigation and corrective action before
advancement resumes.
Safe-SDL applies progression and regression to laboratory safety. PAA generalizes that
transition pattern beyond the laboratory and represents it with reusable contracts.
Malik applies progressive autonomy with safety boundaries and closed-loop verification to
hyperscale network operations. It does not define the evaluator and evidence contracts.
Distillation, active learning, and learning from human preferences show how expensive
judgments can become training data for cheaper evaluation or improved workers. PAA does
not prescribe the training method.