Skip to content

M5 decision package: D-08 learning contract

Decisions

Eligible inputs

Only a server-held feedback.v1 record may originate an experience. The experience type must be one of success, failure, workaround, repeated-question, rework, exception, or pattern-mismatch. Free-form observations without execution provenance remain observations and are not eligible for scheduled candidate creation.

Each feedback record may produce at most one immutable experience. Experiences retain authorization and feedback digests, stable provenance links, privacy class, redaction state, timestamps, and a deterministic deduplication key.

Candidate threshold and deduplication

The default threshold is two experiences backed by two distinct authorization IDs. Repetition within one authorization does not count as independent evidence.

The deduplication key is the SHA-256 digest of normalized scope and recommendation. Observation wording and outcome type remain evidence but do not fragment the candidate. Privacy class is an additional hard partition: evidence of different privacy classes never shares a candidate.

The candidate preserves every experience ID and independent authorization ID. Further matching evidence updates that candidate rather than creating a duplicate.

Cadence and cursor

The default review cadence is seven days. The reviewer reports its current cursor and next due time, rejects an early run, locks review state, and advances atomically over the exact unseen sequence. A stale expected cursor fails closed.

This is an organization cadence default implemented as runtime configuration, not core policy vocabulary.

Privacy

Client-private evidence must arrive through a server-issued redaction envelope; a caller-provided redacted boolean is insufficient. The public MCP intake fails closed for client-private material until such an adapter has issued the envelope. Review reads only privacy classes granted to its control-plane identity. Permission-filtered counts are observable, but filtered content is not exposed. Public review, listing, and proposal-draft operations never reveal client-private candidates.

Each permission profile has an independent cursor and cadence state. A restricted review may advance its own cursor past invisible records without discarding them for a later authorized private reviewer.

Retention

Raw experience content becomes retention-eligible after 180 days. Eligibility is metadata, not automatic deletion. Candidate provenance, digests, review records, and audit meaning remain durable. Any destructive retention mechanism requires separate governed runtime configuration and audit evidence.

Governance boundary

Experiences, reviews, and candidates carry binding: false and policy_change_permitted: false. Candidates additionally carry authorizes_execution: false and explicit proposal, approval, and PR requirements.

The candidate bridge returns inert input for submit_proposal; it has no patch content and cannot submit automatically. The existing governance path remains:

candidate → conductor-proposal issue → Conductor-Approval: approve → PR

Only a merged, valid corpus change becomes live authority.

M4 baseline evidence

M4 supplies the first evaluation evidence, not pre-approved lessons:

Evidence Experience interpretation Candidate effect
V1 guessed the Product workstream and was reverted failure: authoritative state must be read before binding collecting until independently repeated
V2 did not execute JSON Schema, left tax semantics optional, and accepted conflicting writeback failure/rework: declared validation and exact idempotency require executable checks collecting until independently repeated
V3 bound authoritative state, passed independent validation, merged, and wrote back exactly success: the corrected golden-path controls worked evidence only; not a policy promotion

The successful M4 feedback FDBK-F0F434EBB945466D8610E5503A342D12 is an eligible source. Recovery records without a corresponding verified feedback artifact may be represented only after binding them to their durable feedback provenance.

Rejected alternatives

  • One experience creates a candidate: too sensitive to a single anomaly or local workaround.
  • Semantic LLM-only grouping: nondeterministic and able to obscure why evidence was combined.
  • Cross-privacy aggregation: leaks the existence or content of restricted evidence.
  • Automatic proposal submission or corpus edit: collapses the distinction between observation, candidate, approval, and live authority.
  • Automatic deletion at 180 days: can destroy the evidence needed to explain a candidate or audit a decision.