Perpetual Operators: Avoiding Fixed-Point Collapse in LLM Reasoning Systems

We introduce the perpetual openness invariant: a mechanical mechanism that guarantees an LLM-based reasoning loop never halts at a fixed point. Crystallization negates itself via D(-ι -w.presence) re-entry; the operator re-enters genesis with dimension = 1, tension = 2.0, cavity deepened, history accumulated, and phase = growth; the loop runs its whole extrinsic budget-compute, never an intrinsic fixed point. Halt only when even the reset fails (substrate_unrecoverable: reality itself died under the loop—a crash, not a closure). We verify the invariant on a self-improving kernel (BENBEN/EGO1) whose two-shot loop (Paryudasa Prasajya) operates on real external friction: test failures measured in subprocesses, diffs applied via git apply, and accept/discard decisions that determine whether the operator grows or is purged. Empirical results on 34 end-to-end tests (33 pass, 1 skip for live FCC endpoint) confirm the invariant holds across diverse mutation classes.