People now build apps by asking a software assistant, a program that writes code on request, to add or change features over time. Most existing tests hand the assistant a single, fully specified chore and check a specific result, so they miss whether a whole product stays working after many rounds.
This evaluation makes assistants create complete websites and apps from scratch through many requests, then checks only the finished product by automatically clicking through it like a user would. Every feature from every prior request has to survive to the end for the build to count.
Across the large language models tested, the strongest passed 98% of individual checks but delivered a fully working product in only 65% of attempts. That gap shows the hard part is long-horizon reliability, not individual features.