D/E/S: A Tri-Layer Generative Prompt Protocol for Structured Reasoning and Hallucination Suppression in Large Language Models
Abstract
Abstract Large language models generate text token by token, without committing in advance to what a complete answer must contain. This makes them fluent but structurally unreliable: required elements may be omitted, constraints may be ignored, and unsupported claims may appear with unwarranted confidence. Retrieval and chain-of-thought intervene only after generation has already begun; fine-tuning changes the model rather than the prompt. This work evaluates D/E/S, a prompt-level protocol that specifies the components of an answer, their dependency structure, and the epistemic markers for unsupported claims before any text is produced. The study spans six models from four organisations, seven prompt conditions, and 52 tasks across three regimes — framework content, domain-independent reasoning, and adversarial information gaps. Execution quality is assessed through dual LLM-as-judge scoring, inter-judge disagreement analysis, and a five-condition ablation against two published alternatives. The scope of D/E/S is prompt-level discipline: qualifying claims, exposing missing evidence, and enforcing structural completeness. The study maps where such a protocol helps, where it is neutral, and where it fails, providing a boundary-level understanding rather than a mechanism-level claim. Three results. The protocol is followed reliably by the two largest models and unreliably below that tier, and the drop depends on how the structure is encoded rather than on the structure itself. Which encoding works best differs from model to model, and a poor choice performs worse than no protocol at all. Against two published alternatives the protocol improves hallucination suppression consistently across the large models tested; against chain-of-thought the outcome depends on the model. Keywords: large language models, hallucination suppression, prompt architecture, structured generation, D/E/S framework, LLM-as-judge, benchmark evaluation, epistemic calibrationData and full directory structure: https://github.com/gavingu2255-ai/wlm-des-benchmark
