TowerCore: Stacking Descent Systems, Verified in Lean 4 — A Tower That Is Again a Descent System
Year: 2026
Type: preprint
Abstract: A descent system — a step function with a natural-number energy that strictly decreases until it vanishes — carries its guarantees with it: termination within the initial energy, and arrival at a rest state that is never left. This paper asks whether such systems can be stacked: given a list of descent systems acting on one state space, can a single composite system be constructed, uniformly, whose guarantee is inherited rather than re-proved? TowerCore answers yes, under one interface condition: every layer must preserve the rest sets of the layers below it. The tower — defined by towerStep, which runs the lowest unfinished layer, and towerEnergy, which sums the layers' energies, each upper layer measured at the endpoint of the layer below — is proved in Lean 4 to be again a descent system of exactly the base form (towerSys). The kernel's attractor theorem therefore applies to the whole stack at once (tower_halts): every tower, of any height, reaches the joint rest state where every layer's energy is zero, and stays there. What stacking produces is a system of the base form together with its guarantees, not a new physical machine; and the interface is a real condition — arbitrary systems do not stack. The development is self-contained in core Lean, without Mathlib and without sorry; its checked axiom basis is [propext, Quot.sound].
Author: Yamamoto, Takeo
Language: English
Subfield: Software
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7230603835
