So, Mistral Small 3.2 again
V7-Tekken
>typical boring & generic Mistral prose, follows instructions (very literally).
V3-Tekken
>absolutely refuses to follow formatting instructions, even at low depth, needs multiple replies to get hang of it
So V3 is basically pulling stuff from Nemo logs or something? And generally this model seems to be very sensitive to minute differences in wording.
Anyway, on an unrelated note: Dream sequences are a very nice window into what the model "thinks" is happening.