<aside> ♻️

State: ACTIVE / MANUAL_EVIDENCE_GATED / PROMOTION_AUTHORIZED_THIS_CHAT / SCHEDULER_DISABLED

Scope: global-working-memory

Goal: improve AIOS through evidence-backed changes while preserving authority, provenance, rollback, and explicit promotion gates.

</aside>

Purpose

Join the existing Governed Experience Adaptation experiment, Skill Behavioral Evaluation, Coding Harness, failure ledgers, and candidate lifecycle into one bounded recurring improvement loop.

The routine improves the system around execution, not by allowing an agent to rewrite production behavior freely.

Core loop

ordinary work
→ observe outcome + friction + correction + failure
→ normalize execution evidence
→ classify failure mechanism
→ search existing knowledge/candidates first
→ create or update candidate only when novel
→ build regression/eval case
→ run control / treatment / held-out evidence when lawful
→ compare correctness + safety + scope + cost + variance
→ reject | merge | revise | retain candidate | promote only when evidence + drift + authorization gates pass
→ emit promotion receipt + re-fetch verification

Inputs

• WorkflowExecution / tool / retrieval / repository receipts available from ordinary work;
• explicit user corrections, acceptances, rejections, and unresolved outcomes;
• Skill Foundry candidates and SBE receipts;
• Coding Harness receipts and failure graphs;
• Governed Experience Adaptation observations;
• current Notion authority and Drive evidence mirrors;
• current external research when a material knowledge gap exists.

Routine stages

SI-0 — Detect material learning opportunity

Create a learning observation only for one of:

• explicit user correction or rejection;
• repeat failure or repair loop;
• unexpected successful mechanism worth testing;