This study examines backbone extraction not simply as a filtering technique, but as a diagnostic tool to assess the interpretability of archaeological similarity networks. Using a ceramic dataset from the Amuq Valley during the Bronze Age, we evaluate two backbone extraction methods – the Global Statistical Significance (GloSS) filter and the Disparity Filter – in the context of worst-case conditions: limited typological diversity, cumulative survey data, and extensive chronological overlap. In doing so, the paper introduces backbone extraction to archaeological network research as a formal test of relational selectivity, and asks under what data conditions such selectivity can be statistically recovered. Application of the Global Statistical Significance (GloSS) filter produces a null result, indicating that the dense similarity structure is statistically indistinguishable from a randomised null model. The Disparity Filter likewise retains no statistically significant relations: the only edge that survives the procedure links the two sites sharing a rare ceramic class (White Slipped Ware), but it arises from a pair of degree-one nodes — a configuration outside the filter’s domain of applicability rather than a selective tie. Both global and local filters therefore converge on an empty backbone. This convergence is itself informative: the GloSS filter asks whether edges are unusual relative to the entire network, while the Disparity Filter asks whether they are unusual relative to a node’s local neighbourhood, neither recovers selective structure once the redundancy of the underlying data is taken into account. Rather than constituting methodological failure, such negative or minimal results offer the empirical boundaries of relational inference from similarity data. We argue that backbone extraction should be understood as a diagnostic framework that exposes the limits of relational inference, rather than a universally applicable method for network simplification. The approach offers archaeologists a principled safeguard against over-interpreting dense similarity networks, and a transferable criterion for distinguishing genuine relational selectivity from background cultural diffusion.
