"Every number comes from our model, and Claude's job is the words around it" is a clean rule, and the post shows how much work it takes to hold it. The fixed random seed per model version is the detail I liked most, because it means a difference between a plan and a what-if comes from the change and not from a fresh draw. Treating saved plan names as user text that can carry an instruction is also the right call, and an easy one to forget.
At Ascendra Ventures we reached the same boundary from a different direction while building memory for AI coding tools. Memory fixed the repetition, but not the trust: one parser was a single column off and reported numbers that were confidently, uniformly wrong. The lesson we wrote down was to put the model in the middle and never on the edges, and to decide up front which figure the model may propose and which figure plain code must be allowed to throw out. Your MCP server doing no planning math of its own is that boundary in practice.
Have you seen Claude paraphrase a likelihood word in a way that changed its meaning, and if so, did you fix it in the instructions or in the tool output?