The site can generate any sequence of text from its alphabet of 29 characters. This means it contains 29^3200 different works. Let N = log(29)/log(2). It is n

The site can generate any sequence of text from its alphabet of 29 characters. This means it contains 29^3200 different works. Let N = log(29)/log(2). It is not possible to create a compression system where each work on average will use less than N*3200 bits once compressed, assuming each compressed value corresponds to only one work. This stems directly from the pigeonhole principle. What could be possible is to create a compression scheme where say pages consisting of english text can be compressed to smaller (at the expense of pages of random garbage being compressed to longer). This stems from the fact that the entropy (information density) of English text is much lower the 29 character alphabet space can permit. But you cannot on average have the valid English only pages compress to smaller than their entropy either. You can design the system to favor some works over others, but the average compressed length must be equal to or greater than the average entropy of the works. And all you have really done here is invent a weird lookuptable table based compression scheme for english text. All other compression systems are also similarly constrained by the entropy of the data to be compressed, but most other scheme don't require an infeasible large lookup table of basically every possible output.