I think it would be better to require a meaningful percentage (say 1%) of compression rather than an exact count of bytes. Especially while people can ask for arbitrarily large files.
The chance that a randomly generated 1KB file can shrink by 100 bytes is the same as the chance that a randomly generated 100MB file can shrink by 100 bytes.
And that chance is too low to distinguish from zero before the universe dies.
And that chance is too low to distinguish from zero before the universe dies.