Coherence Guard A Deterministic Coherence Constraint for Preventing Hallucinations in Generative Systems

Coherence Guard is a foundational White Paper introducing a deterministic coherence constraint designed to structurally prevent hallucinations in generative systems. Rather than treating hallucinations as accidental errors to be corrected through optimization, learning, or post-hoc filtering, this work starts from a simpler premise: hallucinations are the direct consequence of systems that are required to produce an output even when no coherent answer is possible. The paper formalizes a coherence law, expressed as a simple equation, used not as a quality score but as a condition of admissibility. A response is published only if its coherence exceeds a minimal threshold. Otherwise, the system must explicitly refuse to answer. Refusal is treated as a valid and correct system behavior. The contribution is primarily theoretical. The paper demonstrates that any hallucination necessarily implies insufficient coherence, and that by enforcing this constraint prior to publication, hallucinations become logically impossible, not merely less likely. The White Paper defines: the conceptual framework, the formal properties and guarantees, the explicit limits of the claim. A minimal, non-adaptive reference kernel is provided solely as an executable proof of the constraint, not as the central object of the work.