Coverage decay: when style prompts forget themselves

Stabilising reasoning structure in long LLM outputs Author: Nic Omolabi Format: Technical article / reproducibility protocol Date: May 2026 GitHub repository: PROMETHEUS-74/cogito-coverage-decay Executive summary Large language models do not always fail all at once. They often fail gradually. The first paragraph follows your instructions. The second still resembles the structure you asked for. By the third, the answer begins to loosen. By the fifth, it has become competent, fluent, and generic. This is not a knowledge problem. The model may know the answer. It may even know your preference.