Cool Link
[2601.15714] Even GPT-5.2 Can't Count to Five: The Case for Zero-Error Horizons in Trustworthy LLMs arxiv.org

arxiv.org/abs/2601.15714