— Declare, judge, record, execute — separating authority from the model
People talk about a future in which AI destroys humanity. What if AI develops a sense of self and turns hostile toward humans? That is not the problem this document addresses.
There is, however, something we can do before that moment comes.
While AI is still a system that understands and executes human intent, we can keep it from exercising decision authority on its own. We often attribute the danger of AI to hallucination or faulty reasoning. But there is a more fundamental problem. The model does not stop when it does not know. It fills the blank itself.
All three are the same problem. The model is an engine that fills blanks. This is not a defect. It is trained behavior. Prohibition prompts or better model performance alone will not remove it.
The model distorts the user’s instruction and then executes with confidence. Confidence expands the scope.
A tool will execute as long as its required inputs are satisfied.
This is not a problem of the future. It is happening everywhere, right now.
And solving it does not necessarily require a smarter AI.
Separate both what must be decided and the verdict from the AI.
The model can still do the work. It asks the user, retrieves information, and prepares what is needed for execution.
Rules are declared outside the model as checklists. Code performs the verification and records the verdict. Execution only reads the recorded result.
This takes back the authority the model has been filling in because nobody declared it.
Better model performance increases capability. If decision authority remains with the model, that added capability widens its reach into deciding what was never declared.
If authority sits outside the model, both the decision and the responsibility remain with humans. What this structure does is stop guesses from becoming execution. The goal is not to execute the model’s judgment, but to execute exactly what the user has explicitly declared.
The detailed design is in the specification, and the reference implementation is in the repository.