Build PAA / Implementations
The code carries the detail
These implementations map task boundaries, evidence, and authority changes to real code, and state what each system leaves out.
Normative executable reference
paa-runtime
A Python control plane that validates task declarations, persists evidence, determines promotion eligibility, records authority decisions, and retains event history.
Does not implement: worker execution, evaluator execution, or application-level atomicity for the governed effect. Those remain consumer responsibilities.
Application integration
Scout
Scout embeds paa-runtime in an application that monitors Discord, Farcaster, and Bluesky, evaluates relevance, drafts engagement comments, and presents results to an operator. Outbound content creation and publishing live in a separate application.
Scout declares two PAA tasks: inbound_reply_surfacing, deployed in shadow
mode, and canonical_promotion, currently disabled. Its application database
stores PAA events and resolves autonomy position by exact task, declaration version, and
scope.
Does not demonstrate: an actively enforced production autonomy transition. Scout's grading and feedback loop has produced production-derived evidence, which stays deployment-local and is exported only in redacted, publication-safe form, while its checked-in PAA tasks remain in shadow or disabled deployment. Its publication-safe PAA evidence is explicitly labeled as reference/replay evidence.
Compatible substrate
jig
jig provides bounded execution, typed results, evidence logging, and pluggable graders that map cleanly to parts of PAA.
Does not implement: authority positions, promotion rules, demotion rules, or transitions.
Looking for reusable task shapes rather than code? See Patterns.