If I may, I think it has to do with this... and it's about to get weirder and weirder unless you change the paradigm:

https://safebots.ai/declarative.html

My first-impression is that this site is vaporware generated by someone prompting an LLM to babble about how an LLM could be secured.

At best, perhaps it's an overengineered way to force humans to rigorously validate every action... which the humans are all very hostile to doing.

> [Architecture page] Reasoning is Read-Only: Input=Data, not Instructions

This doesn't make make sense unless the "reasoning" is regular executable code, and then someone still needs to audit the code whenever it changes. The core LLM algorithms simply don't have room for the distinction.

> Complex workflows broken down into encrypted graph segments

Grandiosity++, where does "the architecture" actually end?