Engineering · September 5, 2026
Outcome learning without double-counting
Keep decisions, observations and corrections separate so feedback remains auditable.
Record what could have happened
A decision receipt should bind the authorized candidate set, selected action, context, collection, policy key and selection probability. Record that receipt before executing the action. This preserves the decision-time evidence instead of reconstructing it from whatever happens to look successful later.
Report observations, not guesses
Attach an observed outcome to the receipt when it arrives. Silence, a missing webhook or a delayed customer response does not establish failure. Use stable idempotency identifiers for retries. When an observation was wrong, correct its evidence instead of sending an unrelated second positive or negative result. Test that replay does not inflate the posterior.
Measure improvement outside the feedback loop
Posterior changes demonstrate that feedback was processed; they do not prove customer task-success uplift. Evaluate a controlled baseline and representative held-out tasks, report uncertainty, and watch safety, cost and latency guardrails. Learning must never expand the agent's permissions or replace required human approval.