- Capability purpose
- Turn eligible evidence and interpretation into a durable candidate lesson while keeping later recall and behavior change separately governed.
- Meaningful outcome
- A Learning record exists with evidence, scope, confidence, provenance, lifecycle state, and revision history sufficient for later review or recall.
- Boundaries — includes
- subject and evidence references, source events, lesson statement, originator or model, confidence, applicable context, proposal/confirmation/rejection/supersession state, review timing, and revision provenance.
- Boundaries — excludes
- recording raw observations, detecting patterns, preserving Meaning, changing a profile assertion, creating a standing rule, modifying an agent definition or workflow, executing an adaptation, and selecting retrieval context.
- Terms and concepts
Learning is an evidence-linked proposal for future reasoning. A confirmed Learning remains revisable knowledge rather than self-executing policy.
- Actor
- The owner or an authorized learning process operating within declared scope.
- Command / intent
- Record, confirm, revise, reject, or supersede a lesson that may be useful later.
- Current state
- Eligible evidence and any relevant observation, pattern, Meaning, or prior Learning are available with provenance.
- Policies / invariants
- Machine-generated Learning begins as proposed; the record identifies its evidence, context, confidence, and origin; frequency is not preference; confirmation does not rewrite behavior, policy, skills, workflow, or profile truth; revisions preserve history.
- Transition
- Evaluate the lesson against its evidence, record the bounded statement and applicable context, assign lifecycle state, and link revisions or supersession.
- Result
- A proposed, owner-confirmed, rejected, superseded, or reasoned-not-recorded Learning outcome.
- Events / effects
- Confirmed Learning may be recalled into an authorized context or support a separate proposal, rule, preference, workflow, or agent-definition change.
- Unknowns
- Universal evidence sufficiency, expiry, recall ranking, conflict resolution, and adaptation thresholds are not established.