Reconstruction of Criminal Event Timelines from Property Crime Investigation Documents Based on a Hybrid LLM-Rule Pipeline

Fact-finding in criminal proceedings reconstructs past events into a coherent narrative based on fragmented evidence and testimony, yet remains vulnerable to cognitive confirmation bias. While automating timeline construction using large language models (LLMs) offers decision support, confidentiality-critical local models are prone to temporal hallucinations, inverting causal orders or generating groundless timestamps. To mitigate these vulnerabilities, this study proposes Deep-ARC (Analysis, Reconstruction, and Contradiction detection for criminal timeline construction), a hybrid LLM-rule pipeline that algorithmically regulates model reasoning through legal domain schemas and deterministic rule logic. Deep-ARC injects statutory crime element schemas to categorize event phases and enforces phase-aware boundary clamping grounded on explicit timestamps, systematically resolving sequential contradictions during missing time inference. In proof-of-concept experiments focusing on corporate property crimes (breach of trust and embezzlement) with Gemma-4-31B, Deep-ARC achieved a Micro PMR (exact sequence match ratio) of 0.343 and a Kendall’s Tau (τ) of 0.861, representing an approximate fourfold improvement over the unconstrained baseline and demonstrating higher sequencing consistency than large-scale commercial LLMs. These results indicate that coupling domain-specific structural constraints with deterministic algorithms offers an effective and robust strategy for providing an objective chronological sequence to assist fact-finding without relying exclusively on parameter scaling.