When a Judgment Layer's Self-Reported Fields Lie

A measurement and characterisation study of three judgment layers -- a local non-autoregressive typed-decision model (Laya), a remote typed-decision service (TypeSafe Jev), and a frontier autoregressive model (DeepSeek-V4.1-Flash) -- on a common item set. Four claims are tested; three hold independently (cost is not the binding constraint; access-layer self-reported fields are not trustworthy; capability collapses on tasks requiring the detection of an absence) and the fourth -- that a heterogeneous judge provides incremental coverage over the generator -- is NOT supported in any of three task regimes.