llama 3.3 70b at q4 hallucinates json structure past depth 3, tested 200 examples

tested llama 3.3 70b q4_k_m on structured output generation (nested json with arrays, objects, mixed types) and it completely falls apart past depth 3. success rate on depth 1-2 is ~94%, depth 3 drops to 71%, depth 4+ is maybe 38%. it hallucinates closing brackets, inserts random commas, sometimes just gives up and returns malformed json. switched to q5_k_m and success rate at depth 4+ jumps to 89%. the quantization is definitely breaking something in the json parsing circuit. anyone else seeing this or is it just my sampler settings? temp=0, top_p=0.95, tested on 4090 with 28k context