Why LLMs will always hallucinate: a geometric boundary, not a training bug

I call this constraint RCC (Recursive Collapse Constraints).

It states that any inference system embedded inside a larger manifold —

without access to its internal state, without a stable reference frame,

and without visibility of the container it operates in —

will *necessarily* exhibit hallucination, drift, and planning collapse.

It is not a model bug. It is a geometric constraint of non-central observers.