ReleasedExperiment 0
Two kinds of false "done"
AI models read logs of finished work and reported a status. Models quoted a failing line and still said "done", and rewording the task moved the error instead of removing it.
Sealed · released
Open the record