Digital audio is represented by a series of binary numbers - If you count up to the highest possible number for the number of "bits" (24 bits in this case) and then add one more, it flips to become negative, then as you keep counting it becomes less negative. Similarly if you count as far negative as possible and then keep going, the number will flip to positive.
Here's a simple demonstration counting with 3 bits:
000
001
010
011
100
101
110
111
000 [email protected]!
Now look at this close in zoom of your files: (the upper track is the 16 bit file)
- tracks001.png (20.81 KiB) Viewed 566 times
By default Audacity works with 32 bit sample values, which can handle values over 0dB, but 16 or 24 bit files cannot represent these values correctly.
You could avoid the polarity reversal by deliberately truncating the samples before you export, but obviously that will give you "clipped" (truncated) flat tops and bottoms to the biggest peaks which, while not too noticeable if there's only one or two, is not desirable. A better (and the "right") way to avoid the problem is to Normalize or Amplify the track/mix before you export. I would recommend Amplifying so that the maximum peak is no more than -0.3 dB.