Early in my studies about AI, I learn to get around drifts and hallucinations by replacing text prompts with prompts that are in the style of formal symbolic formats. The end results were more predictable, in that I was able to control the order in which the data was processed (It was another method used to control token usage)