Accepted
2026-07-18
The human maintainer’s APG4 assignment accepts this decision. ChatGPT and Codex execute and review only within that assignment; neither source evidence nor this record expands the authorized phase.
APG3 truthfully stopped before evaluation because its accepted high-rigor comparison required an isolated skill profile, an observable invocation event, and controlled candidate availability and loading. The available environment did not provide those controls without global plugin mutation or new harness machinery. No baseline was run and no candidate was authored.
That result establishes a limitation of the APG3 experiment. It does not show that APG must remain without useful skills until a clean comparison becomes possible. Superpowers remains installed because other repositories use it, but its global presence must not make it the workflow authority for APG itself. APG also needs maturity language that separates provisional usefulness from strong comparative, stable, or decommissioning claims.
Superpowers remains globally installed during APG bootstrap. APG suppresses its workflow authority through repository instructions: Superpowers material is reference evidence unless a human-authorized task explicitly names a specific skill for inspection or comparison.
This is a behavioral repository policy. It does not establish that the plugin is unloaded, absent from model context, or mechanically disabled.
An available source, installed plugin, worker result, report, or familiar workflow is evidence rather than APG process authority. Current APG instructions, accepted APG decisions, and the human-authorized phase define the workflow.
APG may author and retain provisionally useful skills before clean comparative evaluation. Clean baseline-versus-skill evidence is required for strong superiority and stability claims, not for bounded bootstrap authorship.
APG skills use these maturity states:
bootstrap: being shaped and not yet accepted for routine use;provisional: usable within recorded limits after structural, scenario, and
independent review;evaluated: exercised under an explicit evaluation contract with recorded
results and limitations;stable: supported by repeated real-project use and the required transition
review; anddeprecated: retained only for migration, historical compatibility, or
removal.These states describe evidence maturity, not authority, installation, release, or APG’s own distribution license.
Every APG4 skill enters as provisional. Structural checks, public-safe
scenario walkthroughs, independent review, and bounded dogfooding are
acceptable evidence for that state. They establish only that the retained
skill is coherent and usable in the recorded examples; they do not establish
causal, statistical, universal, production, or comparative superiority.
No APG4 skill becomes stable. Stability requires successful use in more than
one real repository and an explicit review after Superpowers no longer governs
the comparison environment.
APG3 remains a truthful blocked experiment. APG4 does not rewrite it as a failed skill, an adequate baseline, or evidence against the candidate. A later phase may retry clean evaluation when Superpowers is no longer required globally or a supported isolated profile exists.
APG records a transition map that routes materially used Superpowers workflows to an APG skill, native Codex capability, repository policy, intentional deferral, or intentional rejection. Useful principles are preserved through APG-native synthesis; fixed templates, universal rituals, and project-specific mechanics are not inherited automatically.
Superpowers decommissioning requires an explicit coverage and rollback gate. APG4 does not satisfy or authorize that gate and does not uninstall, disable, update, or modify Superpowers.
Each skill remains independently removable as one direct-child leaf plus its index, provenance, and evaluation references. A skill that expands authority, leaks private policy, creates unsafe ceremony, or cannot reach a safe provisional state within APG4’s bounded correction limit is omitted or rolled back without introducing a compatibility runtime.
Rejected for bootstrap. It would turn an environment limitation into an indefinite authorship prohibition and prevent practical dogfooding evidence. Strong comparative claims still wait for an appropriate evaluation environment.
Rejected. Other repositories still rely on it, and APG4 has no authority to alter global plugin state.
Rejected. Unsupported profile construction would create fragile local machinery and would not by itself prove matched invocation and loading controls.
Rejected. Its procedures contain source-specific assumptions and ceremony, would duplicate an external workflow, and would not establish APG ownership or validation. Copied or adapted expression would also require preservation of the applicable MIT notice.
Rejected as the bootstrap default. Informal prompts remain useful, but they do not provide independently removable, triggerable procedures or a durable maturity and provenance surface.
Accepted. It permits useful bounded practice while keeping evidence limits, rollback, later comparison, and decommissioning decisions explicit.
evaluated or stable;