Some fixes can be checked without repeating an external generation request. An observed refusal, saved image or delivery record can become reproduction material. The question answered by that check still differs from success of a whole new remote generation. Here I explain how I retain that distinction in the evidence.
An existing answer establishes the state of that answer
An actual Gemini image-edit request ended with provider refusal and no new image. The executor failed to recognize that wording and kept waiting for a nonexistent medium. Testing the classifier did not require resending the same request; the observed response could be reproduced in an owned fixture.
The material represents that response's structure and state. It does not answer whether the account will always refuse, or whether another model or date will produce media. A past response is not a permanent availability declaration. Separating the provider result from the executor's incorrect waiting narrows the code that needs correction.
Keep the pre-fix failure and the opposite case together
The saved pre-fix check expected MEDIA_UNAVAILABLE but received RESULT_NOT_COMPLETE. Final tests include the observed refusal with no new media and a counterexample where the same words accompany an actual new image. The latter must still complete, preventing a wording-only rule from rejecting valid media.
Those unit tests supply defined observations to the executor. They are not a new image-generation experiment against the provider. A passing full regression record does not remove that distinction. Pairing acceptance and refusal cases helps detect a classifier that fixes one delay by broadly blocking legitimate outcomes. Both sides identify the changed behavior more precisely than a larger success count.
Read an existing result before creating another request
Site generation can continue after its observer executor is cancelled for a long wait. A read-only route can inspect the existing response and ready media, then verify downloading that result. Restarting observation does not require restarting generation.
That also does not justify downloading an old image and calling it a new edited result. The Gemini record leaves edited-image and video downloads unexecuted because no new corresponding media existed. Rechecking a stored file, saving an earlier generation and completing a new generation are three different outcomes. The distinction prevents an old success from standing in for a failed new request.
Put the version and action into the verification sentence
A final file-transfer fix passed an independent HTTP receiver check without another actual-service upload on that final build. Earlier live reading and generation remain separate from final-code local evidence. Both are useful, but together they are not one newly repeated end-to-end journey.
For a classification-function question, an existing response and counterexample can support substantial review. Account menus, actual upload acceptance and new video completion require evidence matching those questions. Decide whether reused evidence answers the specific claim before repeating an external request or marking an unobserved stage complete. Precision avoids both unnecessary resubmission and unsupported completion reporting.
An explicitly unperformed test is part of the result
Writing this article compared historical reports, pre-fix results and current test code. It sent no new external generation or refusal request. The public failure lab demonstrates the same separation with in-memory sends and observations and creates no actual media.
If a complete new generation on the latest version has not been checked, retain that empty cell. A later change in availability or purpose calls for reviewing current input and authority before choosing another test. Reusing evidence should sharpen the remaining question while avoiding repetition of completed work. It should never conceal an unperformed stage behind a successful classifier or an old artifact.