The article argues that the fundamental danger of current AI systems is their tendency to fill in missing information rather than stop, leading to execution without explicit instruction. To mitigate this risk, it proposes a framework that removes definition and verdict authority from the model itself.
- The system moves the decision of what must be verified outside the model.
- The verdict on whether verification has occurred is also moved outside the model.
- Rules are declared as checklists, while code performs verification and records the result.
- Execution only proceeds by reading the recorded human-declared result.
This structure ensures that guesses do not become execution, keeping both decision-making authority and responsibility with humans.