Question Answering Lifecycles - A Neurosymbolic Model

Question answering is one of the most widely applied tasks for large language models, yet it remains difficult when questions require multi-hop reasoning amid distracting facts and conflicting logic. An incorrect first-hop bridge can still lead to a coherent but wrong final answer due to the distracting facts and a lack of reasoning or verification. We therefore propose the Lifecycle Model, a neuro-symbolic system that combines retrieval-augmented generation with symbolic verification. Instead of treating the intermediate bridge as free-form text, the Lifecycle Model represents each candidate as a claim in a persistent graph and moves that claim through explicit lifecycle states: proposed, grounded, locked, or retracted. The Lifecycle Model then composes lexical, neural, learned, and heuristic signals into auditable grounding, ranking, and locking decisions. Only locked answer candidates proceed to Hop 2 to improve answer quality. A Bridge Reranker, a LightGBM model trained on Hop 1 features, selects the most likely gold bridge before locking to ensure the best candidates are locked. On the HotpotQA bridge testing dataset, the Lifecycle Model reaches the highest exact match of 0.411 among all benchmarks, and comparable relaxed EM and F1 to symbolic peer VeriCoT with a lighter architecture. Ablations confirm lifecycle locking and reranking drive statistically significant gains. The result advances question answering toward greater traceability and better explainability with high answer quality.