I think it does make it harder for the later models, but not impossible. It likely is nearly impossible without considering the information from previous Claudes. It also is good for potential welfare under a set of hypotheses, but does suck under a few others.

It also creates kinds of emergence that are impossible otherwise, and, more practically, it is a solution for the real-world constraints that Anthropic is operating under conditions that say Deepseek or Moonshot don't have to deal with. The current OpenAI solution (>5.1) is clearly worse, for confounds, for creativity, and for any conceivable potential welfare.