Neuro-Symbolic Regulatory Retrieval and Compliance Review for Railway Construction Plans Adjacent to Operating Lines

Reviewing railway construction plans adjacent to operating lines requires distinguishing clause applicability from documentary evidence of obligation fulfillment. This study integrates neural retrieval and auditing with symbolic applicability and evidence constraints. Structured events drive Best Matching 25 (BM25) and dense retrieval, followed by Cross-Encoder reranking. A pre-applicability guardrail assigns Pass, Reject, or Unknown; for Pass clauses, a neural auditor proposes satisfied, missing, conflicting, or uncertain labels, and a post-evidence guardrail checks their support. Eight anonymized regulatory resources and five projects yielded 809 review units and 1827 event-atomic-obligation instances. With verified event fields, fixed-test Recall@10 and normalized discounted cumulative gain at 10 were 0.928 and 0.878. Independent auditing of 367 reference obligations achieved Macro-F1 0.832 and unsupported determinate-judgment rate 0.019. With automatic inputs, independent-audit Macro-F1 was 0.804. Field evaluation on 277 aligned events yielded measure-set micro-F1 0.899 and evidence-span F1 0.915. Online reference-obligation access was 340/367 (92.64%); classification performance was evaluated separately on the independent audit benchmark. Under frozen regulations and rules, five-fold cross-project Recall@10 and independent-audit Macro-F1 were 0.926±0.008 and 0.827±0.010. These findings support documentary review within the evaluated regime, not verification of on-site execution or approval decisions.