Each lesson is one plain claim of what we found to be true. It names the tests it came from and says where it went. Only lessons from tests on our public record are shown.
The world rules of 64 novels could be turned into 220 general rules that many books share. 94% of these were used by three books or more. A hand check found 27 of 30 rules right.
Changed how we measure It went into a planned feature of our planner.
It came from the step 2026-10-26 · The story ledger
A classifier matched characters' choices to world rules by theme. In Wuthering Heights only 8 of its 19 early matches were right. It matched choices about the living to a rule about ghosts.
Changed how we measure It is not yet in the writer or in a note.
It came from the step 2026-10-26 · The story ledger
A chapter-by-chapter reading with no cap on rules found three to five times more general rules than a capped pass. This held in four novels. The gain grew with the length of the book.
Moved the right way It is not yet in the writer or in a note.
It came from the step 2026-10-26 · The story ledger
With strict matching and merged restatements no rule in three novels shaped many choices before the book named it. The strongest case shaped only three choices.
Changed how we measure It is not yet in the writer or in a note.
It came from the step 2026-10-26 · The story ledger
Rules that shape choices long before a book states them were rare under strict matching. In five novels two were found. Both were quiet duties of a person's role.
No difference found It is not yet in the writer or in a note.
It came from the step 2026-10-26 · The story ledger
Capped and uncapped readings of a book's rules together gave about twice the general rules of the capped one. This was the median over 60 books. In a hand check 25 of 30 rules found only without the cap were real.
Changed how we measure It is not yet in the writer or in a note.
It came from the step 2026-10-26 · The story ledger
A rule list induced again from the merged rules did not label better than the first list. A hand check found 21 of 30 right against 27 of 30. Five misses came from a fault in our design.
No difference found It is not yet in the writer or in a note.
It came from the step 2026-10-27 · The story ledger
A classifier could not reliably find rules a book never states from choices nothing else explains. Only 9 of 20 checked candidates were real. The others were truisms or loose matches on a topic.
No difference found It is not yet in the writer or in a note.
It came from the steps 2026-10-27 · The story ledger2026-10-26 · The story ledger
A rule seldom shaped many choices long before its book named it. With a strict matcher only 2 of 64 books had such a rule. Both were a person's role shown in action before the book put it in words.
Changed how we measure It is not yet in the writer or in a note.
It came from the steps 2026-10-27 · The story ledger2026-10-26 · The story ledger
Sentences from several authors on one topic did not make one scene that read as human. In six collages two model readers scored them about as machine-like as our writing model's own scene. The collages were also hard to follow.
No difference found It is not yet in the writer or in a note.
It came from the step 2026-10-23 · Borrowed pages
Rules for name forms fixed full names said aloud in six borrowed scenes. Errors fell from 21 to 13 and did not reach our bar of 10. The rules did not fix old words of address or minor people that our world does not have.
Moved the right way It is not yet in the writer or in a note.
It came from the step 2026-10-23 · Swap and fit
Our writing model read as less machine-like when it continued a human page than when it wrote from a plan. The gain was smaller than our bar. On 12 whole scenes in our world it still read as machine about half of the time.
Moved the right way It is not yet in the writer or in a note.
It came from the step 2026-10-23 · Borrowed pages
With the scene's plan beside the human page our writing model followed the planned events. It did not read more machine-like than without the plan. This held on 12 scenes.
Moved the right way It is not yet in the writer or in a note.
It came from the step 2026-10-23 · Borrowed pages
Our writing model continued its own text into a second scene in 12 chains. One model reader then read the new scene as machine again. The other still saw a gain over writing from the plan alone.
Moved the wrong way It is not yet in the writer or in a note.
It came from the step 2026-10-23 · Borrowed pages
A borrowed scene chosen by topic alone did not follow from the story so far. Our writing model followed that scene's events and broke the story. Model judges rated its continuity very low in 12 chains.
Moved the wrong way It is not yet in the writer or in a note.
It came from the step 2026-10-23 · Borrowed pages
A fresh human page before each scene kept our writing model's part near human for one model reader. This held over scenes two and three in six chains. The model did not stay near human when it continued its own text with no human page.
Moved the right way It is not yet in the writer or in a note.
It came from the step 2026-10-23 · Borrowed pages
Human pages from other books did not make one story. Each page brought its own people, history and setting. A search for a good fit over scene summaries did not prevent this.
Moved the wrong way It is not yet in the writer or in a note.
It came from the step 2026-10-23 · Borrowed pages
A model judge named the source book of an old human page in 9 to 12 of 12 cases. This was true even with all names changed. The books in the test were well known.
Changed how we measure It is not yet in the writer or in a note.
It came from the step 2026-10-23 · Borrowed pages
A book that model readers did not name from one page was still named from its other pages. In a test on six such books they named the book of every scene with a human page.
No difference found It is not yet in the writer or in a note.
It came from the step 2026-10-23 · Borrowed pages
A borrowed page broke the story even when a plan chose it and an edit merged it to our facts. The edits changed about 8% of the sentences. Scenes without a borrowed page kept the story better.
Moved the wrong way It is not yet in the writer or in a note.
It came from the step 2026-10-23 · Borrowed pages
A human page before the scene did not make our model's part read human when our plan led. This was on eight scenes. The page had helped in an earlier test where the page led and no plan did.
No difference found It is not yet in the writer or in a note.
It came from the step 2026-10-23 · Borrowed pages
A passage by our writing model that read as human did not help its next scene read as human. In six scenes the earlier gain came from the human page and not from the model's own text.
No difference found It is not yet in the writer or in a note.
It came from the step 2026-10-23 · Borrowed pages
An untrained base model did not follow a plan written as a chapter heading. It continued the page in its own voice and ignored the plan. This was on 12 pages.
No difference found It is not yet in the writer or in a note.
It came from the step 2026-10-23 · Trained voices
Our two writing models read about the same when each continued a human page with the scene's plan. In one run on 12 pages both scored about 0.3 to 0.45 on a 0 to 1 machine score.
No difference found It is not yet in the writer or in a note.
It came from the step 2026-10-23 · Prose and render
Texts from the same prompt read in another run moved by up to 0.17 on average. Two texts from one prompt differed by up to 0.67. So a gap between writers read in separate runs can be only noise.
Changed how we measure It is not yet in the writer or in a note.
It came from the steps 2026-10-23 · Prose and render2026-10-23 · Trained voices
A 1.7B base model trained for one epoch on 374 scenes of one author moved toward following a plan. Its machine score fell from 0.92 and 0.98 to 0.82 and 0.88. It was still far from our usual writing models.
No difference found It is not yet in the writer or in a note.
It came from the step 2026-10-24 · Trained voices
An 8B base model trained on 900 scenes of three authors matched their style counts. It still lost the thread of the scene. Model readers still read it as machine.
No difference found It is not yet in the writer or in a note.
It came from the step 2026-10-24 · Trained voices
A model trained on the authors' text did not make our writing model's scenes read human by rewriting them. Each rewrite pass on 24 texts read more like a machine. This held even when the style counts became human.
Moved the wrong way It is not yet in the writer or in a note.
It came from the step 2026-10-24 · Trained voices
A 32B base model read less machine-like than an 8B model trained on the same scenes. It matched our usual writing model with one model reader. It held the scene better but still below our bar.
Moved the right way It is not yet in the writer or in a note.
It came from the step 2026-10-24 · Trained voices
A 31B open model trained on our library's own inputs wrote period prose with human style counts. Two model readers still read it as more machine-like than our usual writing model. It won 0 of 38 blind choices against our usual model.
No difference found It is not yet in the writer or in a note.
It came from the step 2026-10-25 · Trained voices
Adding comic and thriller authors to a romance core did not change how human the trained model's prose read. It also did not change how good it read. These authors made up about a quarter of the training text.
No difference found It is not yet in the writer or in a note.
It came from the step 2026-10-25 · Trained voices
A style classifier named authors well on held-out parts of their own books. On scenes from other books it named one author for every text. This included real scenes by other authors.
Changed how we measure It is not yet in the writer or in a note.
It came from the step 2026-10-25 · Trained voices
The level of detail in a scene plan made little difference to how machine-like our writing model's prose read. Every level read 0.66 to 0.72 on a 0 to 1 machine score. A plan with no words of the story did not read less machine-like.
No difference found It is not yet in the writer or in a note.
It came from the step 2026-10-26 · Prose and render
Rules for name forms needed an exact match of each person to their own list of names. A match on any shared name word gave relatives each other's names. Errors then fell only from 21 to 17.
Unclear It is not yet in the writer or in a note.
It came from the step 2026-10-23 · Swap and fit