llama 3.3 70b q4 - tested coherence on code with function composition vs imperative steps

lol so i got curious if functional composition would hold coherence better than imperative code because it has less mutable state to track. tested same logic in both styles on llama 3.3 70b q4_k_m: function composition (ramda-style): broke at 18.1k tokens imperative with temp vars: broke at 19.4k tokens opposite of what i expected!! my guess is the nested function calls create more stack depth for the model to track even though there's no mutation. anyone else seeing this?