What we believe now

Yes for the parts we use. Scenes now have a closed list of functions and the library is searchable. None of this predicts which books readers love yet.

8 tests · 9 lessons

How it got here

1. 2026-10-26passedCan we lift the rules of each book's world to general rules that any story can use? Passed. The list has 220 general rules in 13 domains. 94% of them are used by three or more books. In the hand check 27 of 30 rules were right. Only 2 rules kept a name after the lift. The classifier left 1.8% of the rules with no general rule.
2. 2026-10-26passedDo novels show most of their rules early, or did our cap on the number of rules make it look so? Our first reading caused the early timing. With the cap 60% to 88% of rules were first shown in the first half. With no cap the four books sat at 61% to 65%. The mean drop was 0.13. This test changed the cap and the reading order together. Thus it does not say which of the two caused the drop. The reading with no cap also found three to five times more general rules.
3. 2026-10-26did not passDo some rules shape what people do before the book names them? Can we find rules that a book never names? Failed. The first matcher linked choices to rules by topic. Only 8 of its 19 early matches were right. A stricter reader that needs a choice to rest on the rule found two wells in five books. All of its early matches were right. The search for unnamed rules found many truisms. At most 10 of 20 candidates were real.
4. 2026-10-26passedAre the rules that only the reading with no cap finds real rules of the world? Passed with 25 of 30. The merged reading gives about twice the general rules of the capped reading for each book. The failures were truisms, one-off remarks and facts of the period. The margin is small. Three of the passes were close.
5. 2026-10-26did not passCan we rebuild the shared list of general rules from the larger merged reading? Failed with 21 of 30. Five rules had no role level because of a fault in our design. Four had a wrong or missing label. The other bars passed. We keep the first list.
6. 2026-10-26no clear answerCan any of our ways to describe a passage find the same move in books by two different authors? The test gave no answer. Two of its checks on the method failed. A beat that we read alone got different events from our full reading in three of four cases. Only 116 of 200 scene rewrites changed enough of the words. The model readers marked 7 of the 980 pairs as twins. No way of describing a passage reached our bar of 80% twins among its closest pairs. On the kept beat rewrites from the books that we kept apart, the embedding trained on stories found the original in 95.9% of cases. Words and typed events each found it in 24.5% of cases. This was a pilot only.
7. 2026-10-27did not passCan we find rules that a book never names, now that we check each candidate against all of its named rules? Failed with 9 of 20. The real ones include refused advances and secrets kept from servants. The others were truisms or loose matches on a topic. Matching by topic is the weak point. We do not use unnamed rules.
8. 2026-10-27passedHow many gravity wells are there in 64 books when a strict reader matches choices to rules? Wells are rare. The reader matched 6,843 choices to rules. 378 of them came before their rule was named. Only 2 rules were wells, in Bleak House and Beverly of Graustark. All 16 early matches were right. Both wells are a person's role shown in action long before the book names it.

← All questions