And you tested this, that the presence of the last full stop flips the outcome with a large SOTA model? Small models are notoriously unreliable and prone to ha

And you tested this, that the presence of the last full stop flips the outcome with a large SOTA model? Small models are notoriously unreliable and prone to hallucination in my experience, so that would not surprise me to be an issue there.