1 | Name what is being evaluated
Identify whether the object is retrieval, generation, routing or page experience. Different objects need different criteria; visual impression and operational reliability are not one conclusion.
2 | Fix inputs and conditions
Retain questions, source versions, user roles and evaluation dates. Results after input changes form a new run rather than being compared indiscriminately with older conditions.
3 | Define acceptable outcomes
Correct answers, inspectable references, correct scope and appropriate hand-offs are distinct requirements. Record why each matters and which user action is affected by a failure.
4 | Classify problems
Separate content errors, version errors, missing references, interaction failures and visual problems. Record reproduction steps and evidence so each issue can be acted upon.
5 | Recheck after changes
After a fix, repeat the relevant checks under the same conditions and examine adjacent flows. Record the result, not merely that a change was made.
6 | Keep the conclusion within its scope
Passing a set of examples means those examples passed under the stated conditions, not that every situation has been proven. A well-scoped conclusion supports further iteration.
