Large language models (LLMs) routinely produce fluent text that is factually wrong, unsupported, or logically inconsistent, a failure mode known as hallucination. Prior work has approached this problem from three largely separate angles: benchmark design (e.g., TruthfulQA), preference-based fine-tuning (e.g., RLHF/InstructGPT), and post-hoc grounding methods (retrieval-augmented generation, chain-of-verification). We argue that these three lines of work are not competing solutions but complementary stages of a single reliability pipeline, and that treating them as such, rather than expecting any one of them to solve hallucination alone, is the more defensible position given the current evidence. We formalize this as a four-stage Draft-Ground-Verify-Revise (DGVR) framework, and we further argue that the Verify stage must address two analytically distinct failure modes: factual unsupportedness (a claim not backed by evidence) and logical invalidity (an inference that does not follow from its premises, even when those premises are individually true). We use this framework to explain why RLHF-aligned models still hallucinate, why retrieval alone is insufficient, and where each existing technique's evidence base actually supports it. We close by identifying the specific open problems, namely claim decomposition, verifier reliability, and calibrated abstention, that would need empirical work to move DGVR from a conceptual framework to a validated system.