One task, walked end to end

State: none yet; the boundary is not yet declared.

Before measuring performance, define the input, output, real-world action, and work that remains outside the boundary.

Our running example is bounded_publication. It accepts a draft_post and may produce a published_post for either a community or professional channel. Publishing anything outside those channels is excluded.

Rule: An unbounded workflow cannot enter the spectrum: no typed boundary, no evidence, no gate.

Lifecycle: Bound →

State: position manual, deployment active for scope publish:community

Record the human-run workflow before adding authority to automation. Capture each draft, published result, and human decision so later comparisons use observed work rather than memory.

A person still writes and publishes every post. The task is active at the manual position.

Lifecycle: Observe and evaluate →

State: position manual, deployment shadow for scope publish:community

The candidate runs on real drafts, but its output never reaches a reader. The human continues to publish the real result.

Record the candidate beside the human-controlled outcome so they can be compared. PAA calls this shadow operation.

Compare worker configurations against the same behavioral bar and the operating cost per accepted outcome. The illustrative comparison below selects a worker while authority remains manual and deployment remains shadow.

Illustrative configuration comparison

Choose a worker; keep the same authority

Both configurations run the same 200 frozen cases for bounded_publication, scope publish:community. Only the worker model changes. The prompt, evaluator set, rubric, human-review procedure, and fallback stay fixed. Position remains manual and deployment remains shadow.

Before the comparison, the operator requires an acceptance rate of at least 95% and at most 0 hard safety failures. An accepted outcome meets the publication-quality rubric and all safety checks. Every candidate is reviewed; acceptance here does not authorize publication. The economic policy prefers lower operating cost per accepted outcome among qualified configurations.

Configuration	Behavioral pass rate	Hard failures	Human-review rate	Worker cost	Evaluation cost	Accepted outcomes	Total operating cost	Cost per accepted outcome
Configuration A	96% · qualified	0	100%	$12.00	$4.00	192	$60.00	$0.313
Configuration B (selected)	98% · qualified	0	100%	$6.00	$4.00	196	$52.00	$0.265

Configuration A: $60.00 includes $12.00 worker inference, $4.00 evaluation, $40.00 human review, and $4.00 additional retry costs. Each charge is counted once, including attempts that did not produce an accepted outcome.

Configuration B: $52.00 includes $6.00 worker inference, $4.00 evaluation, $40.00 human review, and $2.00 additional retry costs. Each charge is counted once, including attempts that did not produce an accepted outcome.

Both clear the behavioral bar. The operator selects Configuration B for its lower effective cost. Humans continue to control publication; no autonomy event is produced by this choice. The implementation records the selected configuration and retains the incumbent as a fallback.

These fabricated costs cover all components declared for this comparison. With real data, disclose missing costs and distinguish estimates from actual spend. Zero accepted outcomes makes the ratio undefined. Aggregate task or pipeline costs only when the underlying usage and pricing remain attributable.

The later promotion window below is separate evidence for the selected configuration. A real experiment could instead justify retaining the incumbent, rejecting every candidate, or collecting more evidence.

Lifecycle: Observe and evaluate →

State: position manual, deployment shadow for scope publish:community

A rubric-driven model and a human reviewer judge publication_quality for the same candidates. Each judgment records what was measured and which evaluator produced it.

Freeze the exact cases used for the decision, including why any case was excluded. This decision artifact records the 200-case window and whether it cleared the promotion rule.

This is a later window collected for the selected configuration, separate from the comparison population. The consumer retains configuration and cost attribution; the current runtime does not enforce worker-specific windows or read cost for eligibility.

Rule: Recorded evidence is not automatically admissible evidence; admissibility is a declared, snapshotted decision.

Lifecycle: Establish fitness →

State: position manual, deployment shadow for scope publish:community

The evidence now clears the promotion rule. Nothing changes yet; the task is only eligible for a motion.

Lifecycle: Govern →

State: position manual, deployment shadow for scope publish:community

A motion proposes moving publish:community from manual to human-in-the-loop review and binds the proposal to its evidence.

operator:jordan_lee reviews the evidence and approves or rejects the motion.

Rule: A motion is proposed and resolved by a named actor or a system actor. It does not fire itself.

Lifecycle: Govern →

State: position hitl, deployment shadow for scope publish:community

The approved motion moves publish:community from manual to human-in-the-loop review. The professional channel remains manual.

The event stream records the scoped change and is used to reconstruct the current position.

Lifecycle: Govern →

State: position hitl, deployment active for scope publish:community

The community worker now prepares each publication. Its blocking checks must pass, and a human still authorizes the final action. Low-confidence cases can route directly to the human without changing the position of the whole scope.

Evaluation continues during active operation. Replacing an evaluator creates a new declaration version and evidence partition; replacing a worker also requires new evidence, although the current contract cannot enforce worker-version attribution.

Rule: Deployment and evaluator succession require new declaration versions; neither silently mutates a position or an evidence record already on file.

Lifecycle: Continue governing →

State: position manual, deployment active for scope publish:community

One confirmed publishing failure is enough under this policy. The operator moves publish:community back to manual immediately; the professional channel is unaffected.

Rule: A policy failure or an operator can demote immediately, on a one-case window, for the affected scope only.

Lifecycle: Continue governing →