An evaluator that only reads code is easy to fool. It will sign off on something that looks right and never runs. So the next thing Sparra learned was to actually exercise the artifact, and for the projects I care most about that means Apple platforms. I wired the evaluator to drive the XcodeBuildMCP CLI: build the app, boot a simulator (auto-discovering one that exists instead of hardcoding a model), launch it, drive the UI, and then read the screenshots back to confirm a UI change is really there.
That multimodal step changed what the judge could catch. It started flagging things a text-only reviewer never would, like an app that letterboxes because it ships without a launch screen. I also gave it the same care a person would: it scaffolds a fresh iOS project with XcodeGen when there is not one, so a greenfield Apple build has somewhere to land. The lesson from the workshop held up here. The contract is only as honest as the check behind it, and the most honest check is running the thing.