A Pluggable and Auditable Search Kernel for Real-World Structural Analysis: Design, Interfaces, and Case Studies
Year: 2026
Type: preprint
Abstract: Many real-world decision problems are NP-hard : power-grid hardening, financial conta gion control, network repair, and frequency assignment. In practice, decision makers do not need provable optimality; they need a usable answer within a predictable time and cost budget, together with an auditable record of how it was obtained. We present a three-layer search kernel that separates a fixed engine from domain-specific problem definitions. A domain implements six methods—build, clone, evaluate a scalar deviation ∆, enumerate actions, apply an action, and validate a state—and the kernel provides budgeting, search, result schemas, and a SHA-256 audit trail. A private engine can be injected behind the same interface, so that intellectual prop erty is not exposed to domain developers. We report five synthetic case studies on commodity hardware: a 12-bus power grid (72.4% deviation reduction, 0.46 s), an 8-bank interbank cascade (100%, 0.05 s), a 12-node brain network (94.7%, 2 ms), a 509-vertex unit-distance graph (a 5-coloring in 69.5 s; non-4-colorability independently verified by SAT and DRAT), and ternary covering codes (n = 4: optimum 9 in 0.09 s; n = 5: 27 in 14 s). We deliberately make no optimality claims for the search layer. Our contribution is a reproducible, low-integration-cost pipeline whose outputs can be checked by a third party, and whose limits are stated explicitly
Authors: Zhang, Baoyuan, Lv, Rui, Qi, Chun
Institutions: Communication University of China, Xi'an Jiaotong University
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7236129440
