I think quality prediction needs to come back in. Right now, all readers see exactly the same information (and thus have identical beliefs), but this has a slight logical inconsistency: if I obtain a book at quality q_i ~ N(q_b, sigma), my beliefs incorporate q_b, not q_i (unless sigma=0, of course).

So let's suppose q_b is not in the information set: how to estimate it? Reintroducing the quality belief is feasible, of course, but I think I can do one better: by combining the quality belief with information from my network neighbours. So essentially a reader would have both a quality posterior and a quality+friend info posterior, the difference being that, once I get a quality draw, I disregard information from friends about that book (I'm deliberately irrational: I trust my own opinions better than that of my friends).

Updating this would probably just be a matter of augmenting the posterior with any unowned books that are owned by friends.

The big change, however, would be that I don't need to use model prediction for books owned by friends: I have a much better signal--the average draw(s) my friends have already obtained.

Thus book selection would involve:

• If at least one friend owns a book, predicted quality is the mean of my friends' quality draws
• Otherwise, I rely on my (augmented) Quality belief.

This might be coupled with another simulation stage: an "information revolution" dictated by low-speed internet, followed by the piracy revolution of high-speed internet.