Multi-Signal Hallucination Risk Estimation: Reference Dependence, Calibration, and Model-Conditional Performance

Reliable hallucination-risk estimation depends not only on the signals available at inference time, but also on how hallucination is operationalized and evaluated. We present an empirical and auditable investigation of multi-signal hallucination-risk estimation, combining white-box uncertainty, black-box consistency, semantic consistency, reference alignment, supervised fusion, and post-hoc calibration. Rather than introducing another detector architecture, we systematically examine which information sources drive hallucination-risk discrimination under explicit information regimes and leakage-controlled evaluation protocols. On an operational archive of 1,890 evaluations spanning nine open-weight models and seven English QA benchmarks, we reserve 20% (378 instances) as an untouched stratified test set. All normalization, model fitting, calibration, and threshold selection use only the 1,512-instance development set. On the locked test set, the full estimator achieves AUROC 0.9646, while reference-assisted features alone achieve 0.9641 and the reference-free regime reaches 0.5850. This near-equivalence identifies reference-based verification as the dominant source of archive-level discrimination, while the reference-free regime is substantially weaker. In a manually labeled, reference-visible evaluation on 315 instances, leakage-controlled fusion across k ∈ {3, 5, 7, 10} folds and 25 repeated resamplings yields AUROC 0.6191, compared with 0.5945 for the untuned score, while reducing Brier loss from 0.2985 to 0.2421. Taken together, these results show that hallucination-risk estimation is strongly conditional on reference information, model behavior, and calibration. We therefore characterize multi-signal hallucination estimation as a reference-conditional, model-dependent, and calibration-aware risk-estimation problem rather than as a universally reliable reference-free detector. Key findings • Reference information dominates archive-level hallucination discrimination. • Reference-free signals provide substantially weaker risk separation on the evaluated QA proxy. • Post-hoc calibration improves probability quality without changing ranking. • Reference-visible human evaluation exposes a substantial construct gap. • Hallucination-risk estimation is conditional on information access and model behavior.