Each atlas shows a novel from before 1931 next to the data that our reading system recorded about it. The notes in the margin show the story at your paragraph. Nothing shows what comes later unless you ask for it.
You will rejoice to hear that no disaster has accompanied the commencement of an enterprise which you have regarded with such evil forebodings. I arrived here yesterday, and my first task is to assure my dear sister of my welfare and increasing confidence in the success of my undertaking.
An atlas shows the reading side of our system. Our writer plans new serials from the parts that this reading finds. These are the questions that a scene raises, who knows what, the threads and the rules of each world. A walk through a book in its atlas shows what the system sees. It also shows where the system reads wrong.
The atlases are also for readers. A classic novel is easier to follow with its map open.
Each panel says what made it. Rules over the parsed text find who is present in a scene: who speaks, who acts, who is acted on and who is spoken to. One language-model call names the members of a group such as “the boys”.
A language model reads the events, the states and the questions of the reader. It does this one time for each book. A classifier checks smaller items: which names are one person, the size of a change and whether a passage shows what a reading claims. Code finds the state that carries forward, the laws and the links between events.
Thus read the model rows as readings to compare with the text. They are not facts. We checked them by hand against the text, scene by scene. They are correct only part of the time.
Scope: hand checks on our first three atlases (A Little Princess, Agnes Grey and Peter Pan), read by an earlier version of the system. We have not yet checked library 1.0 by hand.
These books are famous. The models know them. We changed every name in the chapter ends, and both models still named the book each time. They also predicted much more of the next chapter than they did for obscure books from the same years. Thus the reader in these atlases knows the story.
In our checks no question or imagined answer gave away what comes later. On two obscure novels that no model remembers, the same reading finds about the same mix of questions, answers, turns and foreshadowing. But read the views of the reader as the reading of a reader who knows the book.
We publish the atlas as it is. A walk through a book in this way is how we find these faults.