I programmatically  counted the reasoning word count, rounded it up to the nearest 50, (I.E, 287 words would go to 300, 217 would go to 250), and then I added a note in the system prompt that said:
```
<reasoning:250>
```

and then referenced this in the thinking block.  I think it will work, but I think I need WAY more training data in this format.  Time permitting I light try one if the large reasoning dataset, create the same modifications and try that.