I have been digging causality recently, and I cannot understand what the “do-operator” does. Like I understand how it relates to the notion of intervention, like “what happens to our system when this happens”, but I don’t understand the difference with a simple “conditioned on/knowing” operator.
I feel like the semantic difference is that the “knowing” operator always relies on the probability that X happens, whereas do(X) is more deterministic.
Is my intuition right or somehow right? Or did I get something wrong?
I don’t know if this is dumb to ask, but under PSM, when a system prompt defines multiple overlapping behavioral policies (global rules, context-specific rules, domain rules) that the model must dynamically route between — is that routing itself a persona trait, like the judgment of a skilled professional who knows which hat to wear when? Or is it something more procedural that PSM doesn’t capture?
These two hypotheses predict different failure modes: if routing is persona-mediated, failures look like bad judgment calls; if procedural, failures look like rules dropped regardless of character coherence.