Prajna: Materialization Boundaries and Contract-Driven Planning for Learned Inference

Prajna is an AI-native programming model for learned inference, built around the principle that values produced by probabilistic or learned systems should not be treated as ordinary deterministic values. Prajna introduces two complementary mechanisms. First, a materialization boundary requires model-derived values, represented as beliefs carrying confidence, provenance, and computational cost, to pass through an explicit validation and policy boundary before they can affect durable state or trigger consequential external actions. Second, a contract-driven inference planner automatically selects and composes heterogeneous computational mechanisms—including rules, local classifiers, and large language models—according to declared quality and resource requirements. The paper evaluates Prajna on the Banking77 intent-classification benchmark using both a simulated oracle and Llama 3.3 70B. The experiments compare automated planning with a competent hand-built cascade and investigate cost–accuracy trade-offs, contract infeasibility, selection overfitting, confidence semantics, risk–coverage trade-offs, and deterministic replay. The evaluation also documents implementation failures discovered during real-model testing, including independently occurring infeasible-contract handling defects and a vacuous confidence gate. The broader research objective is a programming model in which learned inference is a first-class computational primitive whose epistemic status, provenance, cost, and permitted effects are explicitly represented, rather than an opaque function call whose output is indistinguishable from a deterministic fact. The paper distinguishes the currently implemented dynamic materialization discipline from a proposed static formulation based on policy-parameterized Verified<T,P> types.