M5 implementation brief: Governed learning loop

Work contract

Field	Value
Milestone	M5
Status	verified
Depends on	Verified M4 golden path
Gaps	G8-G10, G12
Source evidence	M3 feedback plus M4 success and rollback history
Output	Immutable experiences and scheduled, deduplicated, non-binding candidates
Mutation boundary	Learning stores only; never Initiative state, live corpus, approval, execution, or merge

Objective

Turn execution evidence into a reviewable improvement backlog without turning descriptive history or model judgment into policy. A stable-cursor process reviews new, permission-compatible experiences and creates a candidate only after repeated independent evidence.

Work packages

1. #29 — structured experience intake and immutable storage.
2. #30 — stable-cursor review and deterministic candidate deduplication.
3. #31 — privacy, retention, and review-cadence controls.
4. #32 — governance boundary, adversarial evaluation, and independent verification.

Exit conditions

1. Every experience is bound to server-verified non-authorizing feedback.
2. The reviewer consumes only unseen records from an atomic stable cursor.
3. Two distinct authorizations with the same normalized scope and recommendation create one candidate.
4. Privacy classes never cross-deduplicate; client-private input is redacted before review and visibility is permission-filtered.
5. Experience age makes raw evidence retention-eligible but never silently deletes history or changes its meaning.
6. Candidates explicitly cannot authorize execution or change policy.
7. Candidate promotion still requires submit_proposal, Hector's exact approval comment, and a PR.
8. Positive and adversarial paths are independently reproduced.

Non-goals

• Automatically improving prompts, corpus units, or governed assets.
• Treating one successful or failed run as an organization-wide rule.
• Replacing Notion Initiative/workstream state or routine completion.
• Using unredacted private context for background analysis.
• Automatically submitting, approving, merging, or activating a proposal.