A benchmark + accompanying paper introducing a substrate-grounded wrapper for LLMs that grows its own verification capabilities mid-execution while keeping the existing ones verified.
Code (MIT-licensed):
https://github.com/rj45Thompson/sober-ai-complete
Paper PDF (in this repo): sober-ai-wrapper.pdf (9 pages)
GitHub release:
https://github.com/rj45Thompson/sober-ai-complete/releases/tag/v1.1
License (benchmark): CC0 1.0 Universal
License… See the full description on the dataset page:
https://huggingface.co/datasets/RJ45Thompson/bossbench-v1.