llama 3.3 70b loses coherence after 32k context at q4, tested with legal docs

We've been running llama 3.3 70b at q4_k_m on contract analysis (legal docs, typically 25k-45k tokens) and seeing consistent coherence breakdown past 32k context. The failure mode is specific: 1. First 32k tokens: summary quality is excellent, 91% accuracy on clause extraction vs our gold labels 2. 32k-48k tokens: accuracy drops to 67%, model starts hallucinating clauses that don't exist 3. Past 48k: complete breakdown, model will reference sections from earlier in the doc but with wrong details (wrong dates, wrong party names, wrong obligations) Tested on 67 contracts, same pattern every time. The weird thing is this doesn't happen at fp16 or even at q5_k_m - it's specific to q4 quantization. Theory: quantization errors accumulate in the KV cache and corrupt attention past a certain context length. The errors are small enough that they don't matter for short context, but they compound. Anyone else seeing context-dependent quantization breakdown? Also curious if this is specific to llama 3.3 or if other models show the same cliff.