What is the relation between the knowledge base and approved actions?

If the knowledge base is behavioral, rather than propositional, then there is no fundamental difference between the items contained in the knowledge base and the items to be verified. In a sense, both would be behaviors, and the knowledge base could be identified with a behavioral repertoire. The connection between the two could be implemented using a uni-modal autoregressive model on actions.

However, if the knowledge base is propositional, rather than behavioral, but the items to be verified remain behavioral, then there is a disparity in modality. They would be fundamentally different things: propositions and behaviors, despite both being sequences on an abstract level. The disparity could be bridged in one of the following ways.

1. Have a multi-modal model being able to handle and relate both modalities at the same time. Something like Gato.
2. Have a translator from behaviors to textual descriptions, or from behaviors to predicted outcomes to textual descriptions. Then a uni-modal model (e.g. LLM) can relate the textual items.
3. Have a translator pipeline go the other way somehow, from the propositional knowledge base directly to actions. This feels difficult, as the IIE approval would be necessary but not sufficient in identifying relevant behaviors. There's a whole question of context which is not accounted through the KB.