StanceJEV: Typed Probabilistic Stance Detection with Consistency-Verified Decisions

Zero-shot stance detection underpins social monitoring and agent tooling, yet deployed consumers require machine-consumable labels in {Favor, Against, None} rather than free-form text that must still be parsed and re-validated after every call. Recent LLM and small-large collaborative pipelines improve predictive accuracy and query cost, but leave the terminal decision interface free-form even though stance labels already form a closed action set. We formulate zero-shot stance as a typed probabilistic choice over that set, place CoVer-style context reconstruction, batch reasoning, and a consistency gate before emission, and terminate in a Jev/DecisionSOP contract that returns p(y) with an auditable trace. The resulting system, StanceJEV, yields schema-valid labels by construction. On SemEval-2016 Task A (n=1,249) it reaches F_AVG=0.785 at Q_AVG=2.0 with 100% schema-valid typed labels; on VAST zero-shot (n=1,460) it reaches F_macro3=0.833 under the same configuration. Analysis of typed peak probability shows that accuracy rises monotonically across max-p bins, while the consistency score κ characterizes agreement between batch likelihoods and the verifier rather than serving as a calibrated correctness proxy under a single-round budget. These results indicate that evaluating only free-form accuracy leaves the decision interface unsolved: stance monitoring needs a typed terminal contract.