A specification, in Base Layer's terms, is a set of plain-text claims about how a person reasons, communicates and decides. An AI agent reads it before acting for that person. In Base Layer's research, specifications changed how decisions were argued in every situation tested and changed the decision itself in some. That raises a question. Aarik Gulaya, who built Base Layer, puts it this way:
Really, the question becomes, what are choices that only the spec is able to make?
When the model and the specification agree
Part of his answer covers cases where both reach the same conclusion:
in cases where the specification and the base model come to the same conclusion, the only difference then would be that the specification comes with traceability.
Traceability here means each claim in a specification lists the facts it was written from. Why does my AI get me wrong? shows a claim with its evidence.
What is still open
Which choices only a specification can make is still an open question.