Evidence-Calibrated Multi-Signal Hallucination Detection in Large Language Models

Large language models (LLMs) can generate fluent answers that contain unsupported, contradictory, or unverifiable claims,which limits their use in knowledge-intensive and high-stakes settings. Existing detectors tend to focus on one signal, such as retrieved evidence, sampling consistency, uncertainty or internal representations. Thus they suffer differential failure when the evidence is incomplete, the model is confidently wrong or semantically similar answers differ lexically. In this paper, we propose a system Evidence-Calibrated Multi-Signal Hallucination Detection (EC-MHD) to evaluate generated text at the atomic-claim level with five complementary signals, namely evidence entailment, contradiction evidence, semantic entropy, self-consistency and confidence.These signals are fed into a calibrated fusion layer that maps them to a claim-level hallucination risk which is aggregated into a response-level decision with a specific abstention region for ambiguous circumstances. The framework is model-agnostic and can be applied to black-box LLMs with sampling and verbalized confidence, or white-box probability or hidden-state signals if available. We define a reproducible evaluation protocol spanning TruthfulQA, HaluEval, HaluEval 2.0, RAGTruth, FactBench, and processoriented diagnostics, with accuracy, F1, AUROC, AUPRC, calibration error, span-level overlap, and latency as complementary measures. The study contributes a rigorous architecture, mathematical formulation, ablation plan, and failure-analysis procedure without fabricating empirical outcomes; numerical claims are reserved for future implementation and benchmark execution.