Vibeleaderboard
← Glossary
Glossary · AI and agent workflows

Failure class

The kind of failure, named by what the record and the final state show, which in turn decides where the fix belongs.

This module names extraction failures, interpretation failures, unsupported claims, repeated external effects and scope failures. Each is shown by different evidence and is fixed in a different place: obtaining a usable input, a criterion or check about how input is treated, a rule about citing, the workflow's handling of an unknown outcome, or the operation chosen to answer the request. The classes describe what is observable in the record, not the model's internal computation. Naming the class comes before choosing a fix; the proposed fix must address the observed failure and then be checked.

See how the idea fits

What did the run actually do? →
Definition sources