Why a valid setting can still fail
A voice may exist in a configuration and still be incompatible with the selected speech model. Checking that both fields are present misses the relationship between them. If the error appears only at the speech stage, the earlier outline and transcript have already been attempted.
What changed in Open Notebook
We added an early compatibility check for two known Gemini speech models. It considers the default speaker settings and individual speaker overrides before generation and persistence begin. A local catalogue provides the check without another network request. Unknown models and providers retain their existing behavior.
What the local proof establishes
In the invalid configuration scenario, none of the downstream generation stages was attempted. That establishes the failure boundary. It does not attach a dollar value to those avoided attempts or evaluate generated speech.
Where to look in your product
Start with errors that are predictable from information you already have: incompatible options, missing assets, or unsupported combinations. Move those checks ahead of costly work, return an actionable message, and preserve an escape route for configurations your validator does not understand.