# Progressive Autonomy Architecture — full reference

## What PAA is

PAA is a task contract for the space between manual review and unsafe autonomy: a bounded unit of
work whose output is evaluated, gated, logged, and allowed to earn or lose autonomy from the
evidence record. Autonomy should be earned, scoped, and revocable.

## What PAA is not

PAA is not a one-shot autonomous demo, and it does not treat a model's self-reported confidence as
a gate. It is not a maximum-autonomy mandate — many tasks should stay partially governed
permanently. The goal is appropriate autonomy: the level that fits the evidence, the risk, and the
domain.

## Task contract (five orientation groups, eight commitments)

Five orientation groups: boundary, evidence log, evaluator, promotion rule, and demotion. Eight
commitments total — oracle and position policy are evaluator-owned commitments, not peers;
fallback implementation is demotion-owned, not a separate top-level primitive. A task's current
spectrum position is runtime state, not a contract commitment.

- Boundary: typed input in, typed output out; a task is not eligible for the spectrum until its boundary is observable.
- Evidence log: the recorded trail of input, proposal, verdict, and outcome that funds every promotion and demotion decision.
- Evaluator: the verdict-producing mechanism defined by four choices — target (what gets evaluated: input, process, output, or outcome), technique (what produces the verdict: rule, metric, classifier, LLM judge, or human), oracle (what the verdict is checked against: invariant, reference, rubric, human gold, or downstream result), and position policy (how the gate behaves at each spectrum region: blocking, async, or offline). Oracle and position policy belong to the evaluator because they are properties of the verdict producer.
- Promotion rule: the evidence threshold that moves a task toward more autonomy.
- Demotion: two owned commitments — the rule (what breach triggers demotion) and the fallback implementation (the safe lower-autonomy path the task runs when the gate blocks or demotion fires). Fallback belongs inside demotion because it is the target of a demotion, not an independent primitive.

## The autonomy spectrum

assisted -> human-in-the-loop (HITL) -> human-on-the-loop (HOTL) -> autonomous. No automation is
the starting condition, not a level; the spectrum runs as a gradient. Promotion and demotion both
run, in both directions.

## Canonical reference flows and use cases

- PR review agent (/flows/pr-review): Stacked semantic review. Evaluator target: output. Trajectory: HITL → HOTL. Use case: merge safety.
- Refund approval agent (/flows/refund-approval): Selective autonomy classifier. Evaluator target: output. Trajectory: HITL → AUTONOMOUS. Use case: selective autonomy.
- Document routing (/flows/document-routing): Non-generative classification pipeline. Evaluator target: input and output. Trajectory: HITL → AUTONOMOUS. Use case: low-risk classifier autonomy.
- Transaction / trade execution gate (/flows/transaction-gate): Irreversible action gate. Evaluator target: output and outcome. Trajectory: HITL. Use case: irreversible action control.
- Support response agent (/flows/support-response): Evaluator maturity pipeline. Evaluator target: output. Trajectory: HITL → HOTL. Use case: evaluator maturity.

## Illustrative task declaration rendering (non-normative YAML)

The YAML below is illustrative and non-normative. Conformance means meeting all eight commitments,
not matching these specific field names.

```yaml
task: refund_approval
boundary:
  input: refund_request
  output: decision            # approve | escalate
evidence_log:
  - request
  - proposal
  - verdict
  - decision
  - chargeback_signal
  - complaint_signal
  - review
evaluator:
  target: output
  technique: escalation_classifier
  oracle: human_gold          # labeled escalation decisions
  position_policy:
    hitl: blocking            # approve each decision before it executes
    hotl: async               # review batch after execution
    autonomous: offline       # periodic spot checks
promotion_rule:
  metric: recall_on_should_escalate
  threshold: ">= target"
  window: rolling_N_cases
demotion:
  rule:
    condition: chargebacks_or_complaints_exceed_bound
    window: 1                 # point demotion
  fallback: human_review
```

## Prior art and grounding

PAA does not claim to have invented progressive autonomy. SAE J3016 supplies the inherited levels
lineage. Osprey (production safety-critical facility control) instantiates the gated boundary.
Safe-SDL (self-driving laboratories) formalizes evidence-gated advancement and regression. Malik
(hyperscale network operations) independently ports the same pattern. PAA's original contribution
is the evaluator as the governed primitive, the evaluator contract (target, technique, oracle,
position policy), eval-the-eval, and the evaluator maturity curve as the economic engine that makes
gating affordable at volume.

## Canonical routes

- /framework — task model and illustrative declaration rendering
- /manifesto — founding argument
- /evaluators — evaluator taxonomy and maturity curve
- /flows — reference flows index
- /grounding — prior art and grounding
- /contact — contact