An adaptive agent should treat behavior change as a release process. Observed work, positive judgment, candidate training, evaluation, and activation are different events. Collapsing them lets successful execution become an accidental endorsement and lets a new checkpoint change behavior without a promotion decision.
The prior causal spine still applies to operations: source events become working state, state becomes forecasts, forecasts become decisions, and decisions become external effects. Agent authority keeps permission and effect receipts outside the learned model. An effect receipt shows what happened. It does not show whether the agent should imitate that behavior later.
The public Machine extension supplies a concrete learning boundary. Automatic capture is off by default and writes unlabeled observations to a separate file. Only /good creates positive training examples; /bad remains outside supervised fine-tuning. Training defaults to the explicitly judged dataset, and switching into the student model remains a separate command. The trajectory format makes the evidence inspectable across user messages, model output, tool calls, and results.
Co's current synthesis is that improvement needs two adjacent ledgers. The operational ledger records state, decisions, effects, and receipts. The learning ledger records observations, judgments, candidates, evaluations, and promotions. The open question is what evaluation evidence should authorize promotion without reducing every capability, preference, and safety constraint to one reward.