The ability to handle long-context dependencies is a critical challenge for modern language models, especially in tasks that require the retention of information over extended input sequences. Differentiating between latent and positional queries introduces a novel approach to evaluating how well models manage semantic meaning and token order across long texts, offering a more granular perspective on model performance. Through extensive experiments on Mistral, an open-source model, the evaluation focused on tasks designed to isolate latent structure retrieval from positional accuracy. The results revealed that Mistral demonstrated strong capabilities in extracting latent meaning, but struggled with positional accuracy, particularly as sequence length and task complexity increased. A dual-layered evaluation approach enabled a clear distinction between the model's semantic comprehension and its ability to maintain token sequences, uncovering valuable insights into the specific challenges associated with long-context processing. These findings contribute to a deeper understanding of how language models process extended inputs and provide a framework for improving their performance on tasks requiring both semantic and structural coherence.