BERT's tree-metric partially distinguishes constituency depth, with mixed significance

Kennedy applies Hewitt-Manning's frozen, unmodified structural probe (BERT-large-cased, layer 16) out-of-domain to 18,252 synthetic sentences differing only in constituency-tree complement size (vP/TP/CP) with dependency structure held fixed [kennedy-2025-scil] After 78% attrition, a linear mixed-effects model finds significant predicted-distance differences for two of four conditions (SingTP beta=-0.169 p=8.3e-6; DoubCP beta=-0.147 p=0.0004) but not SingCP or DoubTP [kennedy-2025-scil] Marginal R^2=0.27 versus conditional R^2=0.65 shows most variance is driven by which verb is used, not the syntactic condition [kennedy-2025-scil] The author flags an unruled-out finiteness confound (CP complements are finite), so this is partial, confound-uncertain evidence rather than a clean confirmation [kennedy-2025-scil]