however in testing https://raw.githubusercontent.com/alexzhang13/rlm/refs/heads/main/examples/quickstart.py it seems to pick up some system context from nowhere, and has good performance 30% of the time , 60% of the time it's answering some question that wasn't asked . I tried prepending the the example math problem with additional instructions but it only seemed to delay the ghost context from showing up.

(image sometimes works, sometimes does something else entirely. I cannot find references to the ghost context questions that are popping up in the final answer but not sure how to prompt-engineer the "skeletons of the hosted model" away forever with a system prompt) https://cdn.discordapp.com/attachments/1457681408729288784/1457682636913774642/image.png?ex=695ce43f&is=695b92bf&hm=bcf4b74238f4c71dfbcfa3cb94c4779f8ef281adccf0a7cc155663da6fd5ccb0&