I have read http://mattmahoney.net/dc/rationale.html and have two questions, especially about the "How Much Compression can be Achieved?" section. Sorry if these have been discussed earlier.
First, I am not sure my math is right. The rationale says, that no program to date reaches about 1 bit per character, but durilca reaches a compressed size of 127,784,888 bytes (including code+dictionary) vs. 1,000,000,000 bytes uncompressed for enwik9. If I assume 8 bits per byte, that should be a rate of 1.02 bits per character?
The second question is about the fact, that you (the "tester") of course must take into account the dictionary size for the decompression program. On the other hand, the tests that were done with humans (to measure their ability to predict the next character), you did _not_ add any dictionary size. If you would make the test with a first grader, you would get a worse prediction rate. With adults, the rate may approach the cited 0.6 - 1.2 bits per character, but wouldn't you need to add something to account for the information that is "stored" in the brain?