The short answer is that the original encoding of a picture is lost when you re-encode it. So a tool that takes a 70% encoded picture and re-encodes it to 100% actually does the following:
- Decode from 70% to an internal, raw, uncompressed format (similar to .bmp -- basically, it'll be encoded in RGB, probably 8 bits per pixel for each color).
- Encode to 100%
Note two things here: that the encoding assumes nothing about the original image, and that changing the encoding of an image (even from "perfect" RGB to "100% quality" JPEG) is going to introduce scrud -- numerical errors that are for all practical purposes impossible to predict and degrade the image.
So your 70% image isn't perfect -- compared to the original it has numerical scrud. That -- plus a bit more because the conversion isn't perfect -- will be carried forward into the intermediate uncompressed image.
Then, during "compression" to 100% quality any numerical scrud that's in the intermediate uncompressed image will be treated as real image, not as numerical scrud. So when the intermediate raw image -- complete with numerical scrud -- is converted to 100% image quality jpeg, that numerical scrud will be replicated as faithfully as can be managed. Note that, because of the number of steps that the JPEG process goes through, even the best possible job you can do, that results in no compression at all, will still introduce its own scrud.
It actually gets worse than that, but not the way that your question addresses. If you take an image and compress it, then uncompress it, then compress it again, the second compressed image will be lower quality than the first. I honestly don't know if for certain levels of compression you'd reach some minimum of image quality at which the thing would no longer degrade, but I'm sure that if you were to compress to some slightly different image quality each time (i.e., choose a random number between 70% and 80% each time), that the image would just continue to degrade with compression -- even as the file sizes pretty much stayed the same for each compression level.