Naïve Bayes inference for forensic archaeology: Probabilistic artefact-type inference from site context

Crimes against cultural property pose a complex investigative challenge because such cases often involve incomplete recovery, uncertain provenance, mixed-origin assemblages, and fragmentary evidentiary conditions. Under these circumstances, probabilistic reasoning may offer a useful way to structure uncertainty rather than force premature categorical conclusions. This study presents a proof-of-concept Naïve Bayes framework that links broad contextual variables, principally archaeological period and site type, to probabilistic expectations regarding artefact categories. In the present study, the synthetic reference structure and illustrative case applications focus primarily on artefact categories associated with illicit metal-detecting activity and the recovery of metal finds. This reflects the empirical context examined here and should not be understood as restricting the conceptual framework itself to metallic artefacts. The framework is constructed from a literature-informed synthetic dataset and is illustrated through case applications from Fejér County. The article is methodological in scope. It does not claim external validation on an independent real-case corpus, does not compare model performance against specialist archaeological assessment, and is not intended to replace archaeological expertise or function as a standalone evidentiary technique. Instead, it shows how archaeologically informed expectations may be formalised in a transparent, reproducible, and updatable probabilistic form. Within these limits, the framework may support forensic-intelligence and OSINT-based investigative work by helping analysts prioritise contextually plausible artefact categories, structure preliminary hypotheses, and identify which lines of inquiry or online-marketplace signals may warrant closer attention. Internal sensitivity analysis suggests that the model outputs remain relatively stable under moderate perturbations of priors and conditional parameters, while also indicating that the magnitude and interpretability of these outputs depend substantially on the quality and representativeness of the underlying reference data. The proposed approach should therefore be understood as a methodological tool for analytical prioritisation and decision support in the investigation of crimes against cultural property, rather than as a validated predictive or provenance-attribution system.