This preprint presents a comparative analysis of rule-based and LLM approaches for detecting legally actionable red flags in severance letters across Indian and US jurisdictions. We construct a labeled dataset of 150 synthetic severance letters spanning 22 red flag types (12 Indian, 10 US), each grounded in specific statutory provisions. We evaluate a statute-grounded rule-based system (ported from the Lifeboat severance analysis tool, lifeboat-pi.vercel.app) against an LLM baseline (GLM-4) on precision, recall, F1, latency, and statute-citation accuracy. We also propose a hybrid architecture in which the rule engine provides reliable detection while the LLM generates plain-language explanations. Results reveal an unexpected finding: the rule-based system outperforms the LLM on both precision (0.183 vs 0.173) and recall (0.531 vs 0.315), while achieving 500x lower latency and zero statute hallucination. The hybrid system inherits the rule engine's detection superiority while adding LLM-generated explanations. The companion package includes: the 150-letter dataset with ground truth labels, the Python rule engine (ported from Lifeboat's TypeScript), the LLM evaluation harness, the hybrid system implementation, and all results. All code is released under MIT license.
Anamay Anamay Tripathy (Sat,) studied this question.