How AI estimates are made

Most reading levels you see here come straight from a book's catalog record or from a published measure (like Lexile or ATOS). When none of those exist, we estimate a level from the book's description — and we mark those estimates with a faint band and an ⓘ so you always know which is which.

How the estimate works

• A local model. We use qwen3:32b, an AI model that runs on our own hardware — the book descriptions never leave the library's network.
• Three independent reads. We ask the model the same question three separate times.
• It must agree with itself. We only publish an estimate when the three reads agree within about one grade level. If they disagree, we show nothing rather than guess.

Want to see the work? The per-title estimate detail shows each title's three reads, how closely they agreed, and the model's own reasoning.

Estimates are a starting point, not a verdict — a teacher or parent always knows a reader better than a model does.