Skip to main content

Abstract

Boolean Satisfiability (SAT) is a central problem in theoretical computer science and a cornerstone of modern industrial reasoning systems. While state-of-the-art solvers based on Conflict-Driven Clause Learning (CDCL) achieve impressive results on many benchmarks, they remain fundamentally black-box systems and exhibit severe limitations on globally constrained instances. This paper presents a Glass-Box resolution engine that treats SAT resolution as a process of structural closure rather than heuristic search. On a frozen class of non-local SAT instances (global XOR/parity constraints encoded as CNF), this approach robustly closes all instances while producing complete, auditable resolution traces. These results do not claim to solve SAT in general, but establish that a meaningful subclass of NP problems is already better handled by a structural, auditable reasoning paradigm.

Key Results

100% closure

All instances closed across all tested sizes (400–1600 variables), both SAT and UNSAT.

Glass-Box artifacts

Every execution produces a verifiable certificate, a complete event trace, and deterministic replay data.

Robust under global constraints

The target class (global XOR-CNF) is specifically chosen to defeat local reasoning strategies.

What This Paper Demonstrates

The paper establishes three properties on the target problem class:
1

Structural resolution

A class of non-local SAT instances can be resolved through representational closure rather than heuristic search.
2

Process-level auditability

The reasoning process itself is an observable, replayable mathematical object — not just the final answer.
3

Comparison with established solvers

On this specific class, the Glass-Box engine closes 100% of instances where a modern CDCL solver (Kissat) times out on all satisfiable cases.

Experimental Summary

Size (variables)R3-MRM Closure RateCDCL Baseline Closure Rate
400100%40% (UNSAT only)
800100%40% (UNSAT only)
1600100%40% (UNSAT only)
The CDCL baseline resolves UNSAT instances rapidly but systematically times out on all SAT instances at every tested size. The full per-instance results, including step counts, contraction behavior, and runtime data, are available in Appendix A of the paper.

Scope and Non-Claims

This paper is intentionally conservative in scope. To preserve scientific credibility, the following are explicitly not claimed:
  • That SAT is easy in general.
  • That NP problems admit efficient solutions universally.
  • That this approach replaces CDCL on all instance classes.
  • Any implication toward resolving P = NP.
The target class (global XOR constraints encoded as CNF) serves as a structural stress test. Results must be understood in relation to this structural specificity, not as a blanket claim over all SAT.

Why This Matters

Beyond SAT, this work demonstrates that it is possible to resolve non-trivial NP-class problems while simultaneously producing artifacts suitable for independent audit. This challenges the widespread assumption that high-performance reasoning systems must necessarily be opaque.
The same discipline of proof — explicit traces, convergence measurement, and verifiable certificates — is intended to transfer to broader Representation Model instantiations, including language and multi-domain reasoning.

Relation to R3-MRM

This paper constitutes the first published validation of the R3-MRM system. It demonstrates that the RM architectural class, instantiated in a mathematical domain and executed on the QDE computational core, produces measurable and independently verifiable results.

Citation

Mazzoni, S. (2026). Structural Resolution of Non-Local SAT Instances: Glass-Box Closure for a Subclass of NP Problems. RCUBEAI Research.
All instances, logs, and public verification traces are available upon request or via a controlled API under the same experimental protocol.