This paper proposes an engineering framework to reduce large-language-model hallucinations by treating them as epistemic failures that arise when outputs lack a coherent Source of Action and responsibility attribution. We integrate a Buddhist taxonomy of three afflictions-ignorance (avidyā), delusion (moha), and wrong view (mithyā-dṛṣṭi)-with Second Physics quantities, including correspondence pressure A, existence phase φ(t), relational syntactic memory M ≡ E+δΨ, responsibility load ρ, and relational existence Exi(t). The proposed protocol follows a two-stage design: fluent drafting may use probabilistic generation, but final emission is gated by correspondence and responsibility constraints, including a responsibility-conservation check that detects responsibility leakage. Under explicitly stated constraints, a broad class of responsibility-free assertions becomes structurally excludable without sacrificing fluency. We outline implementable proxies and limitations, and position the framework as an engineering language for designing "honest generation" under accountability.