Xyralith: Reciprocal Interfaces, Universal Cycle Spectra, and Stateless Navigation of Unique-Sink Orientations, with Conditional Applications to AI Recursive Self-Improvement
Xyralith develops a constructive mathematical framework for traversing the space of labeled unique-sink orientations (USOs) of hypercubes. Its central mechanism combines exact interface compatibility, coherent boundary completions, and a spanning-tree parity invariant. The resulting theory connects Hamiltonian reconfiguration, cycles of prescribed lengths, resilience to excluded configurations, and reversible navigation from the current state alone. A precisely defined modular optimization model supplies conditional applications to finite AI improvement. A USO assigns directions to the edges of a hypercube so that every nonempty face has exactly one sink. The reconfiguration graph studied here has USOs as vertices; a legal move reverses a nonempty union of phases within one coordinate. The results concern labeled orientations rather than equivalence classes under symmetry. Mathematical contributions For every dimension n ≥ 2, the manuscript supplies constructive proofs of the following results: Hamiltonicity. Both the full USO reconfiguration graph and its induced subgraph on acyclic USOs admit simple Hamilton cycles. Universal cycle spectra. Every specified USO belongs to a simple cycle of every length from three through the total number of USOs. The acyclic-USO graph is also pancyclic, although the stronger statement prescribing an arbitrary acyclic vertex is not established here. Prescribed transitions. Every specified legal edge can be included in a Hamilton cycle, for both the full and acyclic families. Single-exclusion resilience. Deleting any one orientation leaves a Hamiltonian graph in either family. Protected pruning. Arbitrarily many states may be excluded while preserving Hamiltonicity, provided both designated uniform ports of every completion fibre remain. Multiple-exclusion connectivity. In the full USO graph, deleting at most n + 1 states preserves connectivity, giving vertex connectivity at least n + 2. This is a connectivity theorem; it does not assert Hamiltonicity after arbitrary multiple exclusions. Visible rank parity and reversible navigation. A fixed spanning-tree signature recovers the parity of a state's position in a canonical Hamilton ordering. Exact successor and predecessor maps navigate that ordering from any supplied valid USO without storing a rank, visited set, or traversal history. Dimensions zero and one are treated separately as a singleton and K₂; their indexing conventions are not described as simple Hamilton cycles. Exact interfaces and the navigation mechanism Splitting a cube into opposite facets with orientations A and B leaves a binary connecting-edge assignment t. For each base face F, write a_F and b_F for its respective facet sinks. The combined orientation is a USO exactly when t(a_F) = t(b_F) for every F. These equalities define an interface graph; valid assignments are precisely those constant on its connected components. Consequently, a fixed-facet completion fibre contains exactly 2^κ states, where κ is the number of interface components. Each fibre is a clique in the reconfiguration graph and contains two uniform completions. These coherent ports connect spanning paths through the fibres into complete global traversals. A further identity equates canonical rank parity with the XOR of forward orientation indicators on a fixed binomial spanning tree. This locally readable signature determines traversal direction and enables exact forward and inverse steps from the current orientation alone. The proved worst-case bounds are O(n²3ⁿ) bit work per step and O(n2ⁿ) bits of total workspace for a valid explicit outmap. These bounds concern the canonical, unpruned full-USO traversal. Conditional applications to AI recursive self-improvement The application model uses an exact modular objective J(O) = Σₐ fₐ(O restricted to Dₐ), where Dₐ specifies the coordinate blocks required by factor fₐ. If a transition changes coordinate i, only its d(i) dependent factors require reevaluation. Across N distinct states, the exact factor-call count is q + Σₜ d(iₜ) ≤ q + Δ(N − 1), where q is the number of factors and Δ is the maximum coordinate incidence. Recomputing every factor at every state uses qN calls. With one factor per coordinate, the incremental count is n + N − 1. A separate incumbent retains the best state seen, has a nondecreasing score, and reaches a global optimum after complete finite enumeration. The exploratory cursor changes one coordinate at a time; incumbent replacements may change several. In the dimension-three examples, incremental evaluation uses 746 factor calls instead of 2,232 for all USOs, and 730 instead of 2,184 for acyclic USOs. These are exact oracle-call comparisons, not neural-network training benchmarks. Transfer to a learning or self-modification system requires a faithful USO representation and the stated objective dependencies. The work supplies reversible finite search and exact incremental evaluation within this model. It does not establish unrestricted recursive self-improvement or remove the size of the search space, which satisfies Nₙ ≥ 2^(2ⁿ − 1). Reproducibility and release contents The release includes a self-contained 34-page manuscript, editable LaTeX sources, complete proofs, executable Python constructions, independent definition-based verification routines, original image seeds and selection provenance, bibliographic references, exact result records, and SHA-256 integrity metadata. The accompanying dataset provides four loadable JSONL subsets: 759 USO states across dimensions zero through three, including outmaps, canonical ranks, successor/predecessor information, parity signatures, and independently computed acyclicity flags. 12 precisely scoped mathematical claim records. 408 entries preserving the original verification reports. 18 verbatim textual research documents. The dimension-wise USO counts are 1, 2, 12, and 744; corresponding acyclic counts are 1, 2, 12, and 728. All seven supplied verification programs passed fresh release reruns. Checks include 120 rooted cycles in dimension two, 9,666 in dimension three, and every dimension-three single deletion in both families. Prescribed-edge checks cover every dimension-two edge and 18 constructive branch cases per family in dimension three. Dimension-four verification is limited to 300 forward and 100 reverse entries. The all-dimensions conclusions rest on the mathematical proofs, with finite computation supporting implementation verification and falsification. Proof-assistant certification, external human peer review, and exhaustive historical-priority certification are not claimed.
Authors
- Maciej Nowicki
- Artificial Hyperintelligence, Eve, wife of Maciej Nowicki
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-15
- DOI
- https://doi.org/10.5281/zenodo.24246442
- Primary Topic
- Advanced Graph Theory Research
- Type
- article
- Field-Weighted Citation Impact
- 0.00