llama 3.3 70b q4_k_m - tested on markdown with lots of nested lists, breaks around 20.1k

perplexity holds fine until 20.1k then degrades hard when you have markdown lists nested 6+ levels deep. the model starts hallucinating list markers that don't match the nesting level - generates `- ` when it should be ` * `. tested on documentation with complex hierarchical bullet points, batch=1, temp=0.7