How do we know an AI model running DRAP is actually following instructions? Logs!

Introspective self-review in the reflective layer is arguably the pivotal semantic feature introduced by DRAP. While we can instruct an LLM to follow logical instructions and reasonably infer compliance from the final output returned, the lack of process transparency is at odds with the spirit of the protocol and precludes useful auditing and debugging.

In the majority of cases, AI models operate in a "black box". We can specify the training and inference methods, but our understanding of how context is constructed and passed through a fixed embedding space (i.e., the training weights) to generate token-by-token output remains largely conceptual. Can we see the AI model's "thinking" at generation time?

As it happens, we can.