Keep the lesson for the next incident
Classify certainty failures
Group incidents by missing evidence, stale source, ignored alternative, unsupported inference, forced output format, copied certainty, unclear claim type, or absent confidence standard. Keep one claim and decision consequence for each class. Distinguish bad wording from a genuinely unresolved question.
Keep matched exemplars
Retain cases pairing evidence strength with appropriate cautious, supported, strong, and provisional language. Include contradictory and non-discriminating examples. Annotate claim type, source limits, alternatives, reviewer action, and why a numerical confidence label would or would not be justified.
Review recurrence and ownership
Assign owners for evidence standards, task wording, result review, and policy choices as needed. Sample material agent outputs on a defined cadence and track unsupported certainty by class. Reopen investigation for repeated overstatement, changed evidence coverage, or confidence labels used outside their contract.
What to carry forward
Close the learning record with certainty-failure classes, matched exemplars, wording standards, owners, and a trigger for unsupported or out-of-scope confidence. Preserve honest provisional language as valid behavior. The durable result is reviewable certainty tied to evidence, not a demand for artificial precision.
Keep the decision with the work.
Use a Work Item in Aglet to record the problem, the evidence you have, and the next decision. Add an owner and priority, then keep updates in the discussion so the next person can follow the reasoning.
Create an account See the product workflow