llama 3.3 70b q4_k_m perplexity degrades faster on yaml with anchors vs plain yaml

We tested same config file in two formats - one with yaml anchors and references, one with everything expanded. Plain yaml holds coherence to around 26k tokens, anchored version breaks at 21k. Perplexity at 20k: plain yaml is 7.8, anchored is 11.2. Same content, just different representation. Is this tokenizer thing or model actually struggling with reference resolution? We need to decide if we should convert all our configs before processing.