This paper presents empirical evidence for a novel approach to large language model (LLM) alignment and hallucination reduction. A compact system prompt — the Foundation Prompt — anchored by a logically self-grounding statement restructures model reasoning from the ground up rather than constraining it from outside. Three independent studies are reported: a 60-run five-platform ablation study (600 individual metric assessments), a 10-question cross-platform comparative study, and a 16-question cross-domain replication spanning clinical medicine, physics, mathematics, law, regulatory science, social science, finance, philosophy, and frontier physics. Key findings: structural hallucination is eliminated entirely on three of five major reasoning-native platforms and reduced by 48–100% across all reasoning-native platforms; composite hallucination risk is reduced by 50–60% per platform; a novel confabulation anchor mechanism is identified, characterising the specific conditions under which hard (Category 1) hallucination is structurally most likely in unaligned LLMs and enabling advance detection of highest-risk question types; and a recursive coherence property is demonstrated, whereby aligned outputs are higher-quality inputs for subsequent reasoning steps, creating compounding improvement trajectories with direct implications for agentic AI safety and recursive self-improvement research. A four-category hallucination taxonomy is introduced — distinguishing hard hallucination, precision hallucination, structural hallucination, and omission hallucination — providing a more granular measurement framework than existing binary approaches. The Foundation Prompt requires no architectural changes, no model retraining, and no fine-tuning. It operates on currently deployed models via the system prompt interface. The prompt text and full methodology are available to researchers on request. Underlying intellectual property is blockchain-timestamped from January 2026.
Steven Daw (Wed,) studied this question.