Controlled Evolution Governance
- Operate only on structured observation facts, statistics, evidence references, assessments, and proposals.
- Every proposal keeps
execution_authorization=NONE.
ACCEPT permits creation of a separate implementation task. It does not grant file, Git, deployment, production, or data-write authority.- Aggregate only exact
project_id + repo_fingerprint matches. Deduplicate by event_id, then aggregate by task_id.
- Terminal outcomes are only
PASS/BLOCKED/FAILED/CANCELLED/PARTIAL/UNKNOWN; generic status fields cannot infer outcome.
- Hooks retain minimal structured metadata and never raw prompts, full answers, source bodies, patches, tokens, cookies, API keys, or credentials.
- Reviewers, explorers, and workers share one root budget. Registered Reviewers start from Luna and score valid evidence before selecting up to Astra High; other roles retain the original four profiles. Never read, infer, store, or export host model identity.
- Pass the health gate before analysis. Project automation requires explicit opt-in. Keep NO_CHANGE, WAITING_FOR_TASKS and COOLDOWN quiet; report only meaningful changes, new candidates or actionable failures.
- New proposals freeze testable hypotheses. Separate implementation validation from benefit proof and revalidate references, windows, independent samples and quality guardrails on replay. Confirmed causes generate pending regression candidates without automatic implementation or cross-project promotion.
Lifecycle Hooks -> TaskOutcomeEvent V3 -> task aggregation
-> project isolation -> Snapshot -> Assessment -> Proposal
-> human decision -> separate implementation task -> normal validation and delivery gates
Use Luna for mechanical aggregation and Terra only for material cross-task conflicts or high-risk policy decisions. Higher model effort cannot repair poor evidence quality.
DelegationBudget calibration consumes only parent-finalized, project-bound samples with approved-profile attribution and cost-basis units. Insufficient adjacent-tier samples require no change, and every proposal retains execution_authorization=NONE.
1---2name: controlled-evolution-governance-23description: Use only for cross-task retrospectives, observation governance, model-cost routing, reviewer calibration, routing drift, proposal lifecycle, project isolation, and assistant version governance.4---56# Controlled Evolution Governance781. Operate only on structured observation facts, statistics, evidence references, assessments, and proposals.92. Every proposal keeps `execution_authorization=NONE`.103. `ACCEPT` permits creation of a separate implementation task. It does not grant file, Git, deployment, production, or data-write authority.114. Aggregate only exact `project_id + repo_fingerprint` matches. Deduplicate by `event_id`, then aggregate by `task_id`.125. Terminal outcomes are only `PASS/BLOCKED/FAILED/CANCELLED/PARTIAL/UNKNOWN`; generic status fields cannot infer outcome.136. Hooks retain minimal structured metadata and never raw prompts, full answers, source bodies, patches, tokens, cookies, API keys, or credentials.147. Reviewers, explorers, and workers share one root budget. Registered Reviewers start from Luna and score valid evidence before selecting up to Astra High; other roles retain the original four profiles. Never read, infer, store, or export host model identity.158. Pass the health gate before analysis. Project automation requires explicit opt-in. Keep NO_CHANGE, WAITING_FOR_TASKS and COOLDOWN quiet; report only meaningful changes, new candidates or actionable failures.169. New proposals freeze testable hypotheses. Separate implementation validation from benefit proof and revalidate references, windows, independent samples and quality guardrails on replay. Confirmed causes generate pending regression candidates without automatic implementation or cross-project promotion.1718```text19Lifecycle Hooks -> TaskOutcomeEvent V3 -> task aggregation20-> project isolation -> Snapshot -> Assessment -> Proposal21-> human decision -> separate implementation task -> normal validation and delivery gates22```2324Use Luna for mechanical aggregation and Terra only for material cross-task conflicts or high-risk policy decisions. Higher model effort cannot repair poor evidence quality.2526DelegationBudget calibration consumes only parent-finalized, project-bound samples with approved-profile attribution and cost-basis units. Insufficient adjacent-tier samples require no change, and every proposal retains `execution_authorization=NONE`.
JimmyVGDY