Why reading-time prediction needed per-user data

A words-per-minute constant is the obvious first approach to estimating how long a piece of content will take someone to read. It's also wrong in a way that matters: reading speed varies enormously between people, and even for the same person, it varies by content type, familiarity, and time of day.

In the Personalized Reading System, the model instead treats reading time as a small regression problem. Content length is still the primary input, but it's combined with a user's own historical reading pace — how long they've actually taken on similar content before — and engagement signals like time spent per page and completion rate.

The practical effect is that the same 20-page chapter gets a different estimate for two different users, based on how each of them has actually read in the past. That's a small change in modeling approach, but it's the difference between a number that's technically present and one that's actually useful.

There's a second benefit that wasn't obvious until later: because the model relies on behavioral history rather than only content metadata, it also naturally feeds into the recommendation logic — a user's reading patterns inform what gets suggested next, not just how long it will take.