← Glossary
Glossary · AI and agent workflows
Completion evidence
What will be checked to accept the work, named before the work starts so that acceptance does not depend on the final message.
Completion evidence names things other than the model's final message: the output file read by a person, the action record showing each input was read, a receipt from the service that performed an operation, a calculation redone by hand. It exists because a brief cannot guarantee correctness, and it is checked against the record and the output after the loop ends.