01
Research object
The model should choose what to observe next when information is incomplete. The work is defined as a measurable computational problem rather than a brand category.
02
Method
We use locked tasks, paired cases, hidden worlds, formal or executable evaluators where possible, and explicit failure classes. The metric is chosen to match the scientific question, not to rescue a preferred architecture.
03
What would change our mind
The program is falsifiable by construction. If a simpler token-centric system achieves equivalent state stability, discovery yield, or decision quality under matched information and resources, the architecture must be revised.