The question
Can a personal system powered by language models become more dependable over time rather than accumulating invisible failure?
The experiment focuses on observability, graceful degradation, and making the system easy to correct when its model of reality drifts.