2026-10-29 · The lift · Reading a novel
What we tried
We tested whether recent changes to our genre map change the labels on books. A model labelled events and characters twice each with old and new options on crime, caper, romance and control books.
What we learned
All three tests failed. Most changes altered no question that a book gets and so changed no label. The fix for a dark romance changed which types we chose but did not cut misfits or win over two model judges.
1 test ended in this step: 1 did not pass.
2026-10-29 · The lab's record · Several theories
What we tried
Two model readers read ten episodes of our own serial as listeners and noted what loses them. We compared their notes with the same reading of two human-written romance serials.
What we learned
No fault in our serial reached twice its rate in the human serials. The two readers noticed different kinds of fault. Continuity slips came closest and our check had flagged most of them before we published.
1 test ended in this step: 1 no clear answer.
2026-10-28 · Planning · Several theories
What we tried
We dropped one test of a new plan store because no likely result could meet its bar. We studied why the romance in our serial felt slow and re-planned it under a measured pace.
What we learned
The serial reuses its own phrases. Its romance stops at a touch where the romances we compared it with move on to admitted feeling. Re-planning showed that the pace must come from the chapters because scene plans only put romance labels on deal scenes.
16 tests ended in this step: 5 passed · 10 did not pass · 1 dropped.
2026-10-28 · The lift · Reading a novel
What we tried
We tested two ways to find beats by different authors that make the same story move. One matched summary words and one matched typed story events alone.
What we learned
We dropped both before review because a better design replaced them. Its readers found almost no true matches. Neither method works: words match under 1% of beats and events match barely more than random events do.
2 tests ended in this step: 2 dropped.
2026-10-27 · Planning · Several theories
What we tried
We tested upward plan revision, story links in the prompt of our writer and a register of variables for plan edits. We also planned four tests on classic novels: transposition, blends, payoff threads and reader knowledge gaps.
What we learned
Plans with no links beat plans with correct links. Our extraction failed twice and we dropped the four tests on classic novels. Locked plan levels blocked upward revision and most register edits so neither test gave an answer.
7 tests ended in this step: 2 did not pass · 1 no clear answer · 4 dropped.
Drew on
2026-10-27 · The story ledger · Reading a novel
What we tried
We ran four tests on our reading of 64 public-domain classic novels. We rebuilt the shared rule list, hunted unstated rules, sought early hidden rules and lifted story objects to general levels.
What we learned
Three tests failed hand checks. These were the rebuilt rule list, the unstated rules and the typing of lifted objects. Only 2 of 64 books have a rule that shapes many choices before the book names it.
4 tests ended in this step: 1 passed · 3 did not pass.
Drew on
What we kept
2026-10-27 · The lift · Reading a novel
What we tried
We built the story rules, tensions, arcs and plot threads of three classic novels with code and one model pass. We scored them against reference lists that two models made from each text.
What we learned
The test failed. The code found almost all plot threads but put the main tension first in no book. It split one pressure into several tensions and gave the top rank to the wrong one.
1 test ended in this step: 1 did not pass.
Drew on
2026-10-27 · Scene index and search · Several theories
What we tried
We filtered our scene search by what each scene needs from the story and what it gives. On 17 story problems we tested it against the plain search and a filter for the wrong problem.
What we learned
The filter failed. Its picks fixed no more problems than the plain search. A filter for the wrong problem did just as well and we keep the plain search alone.
1 test ended in this step: 1 did not pass.
Drew on
2026-10-26 · Trained voices · Prose that reads human
What we tried
We trained an open model to write the text of the authors from plans with no story words. We wanted to know if it reads less like a machine than the same model trained on our usual beat summaries.
What we learned
We stopped the training after 32 minutes to move to more promising work. The model generated no text and the test gives no answer. Earlier tests suggest a wrong approach because they found no twin scenes across authors to train on.
1 test ended in this step: 1 dropped.
2026-10-26 · Prose and render · Prose that reads human
What we tried
Our writing model wrote 12 held-out scenes from three plan levels with no training. The levels were rich plans, our beat summaries and plans with no story words.
What we learned
The test failed because the no-words plan read slightly more machine-like than our summaries and the model followed it less. Every plan level scored 0.66 to 0.72 on machine reading against 0.03 for the real text. So the writer is the problem.
1 test ended in this step: 1 did not pass.
What we kept
2026-10-26 · The story ledger · Reading a novel
What we tried
We lifted the world rules of 64 old novels to 220 general rules and read rules chapter by chapter with no cap. We looked for rules that shape choices before a book names them.
What we learned
The general rules are widely shared and a full reading with no cap finds about twice the rules of a capped one. We cannot tell why our first reading made rules seem to come early because it changed two things at once. Our search for hidden rules failed.
5 tests ended in this step: 4 passed · 1 did not pass.
What we kept
2026-10-26 · The lift · Reading a novel
What we tried
We built a test that checks whether a method finds the same story move in different words. We measured four methods with it: plain words, typed events and two text embeddings.
What we learned
The test failed two of its own checks and we must run it again. A narrative-tuned embedding found rewritten passages almost always but words or typed events found them one time in four. Model readers judged almost no closest match from another author as the same move.
1 test ended in this step: 1 no clear answer.
Drew on
2026-10-26 · Scene index and search · Several theories
What we tried
We replaced the word search in the first stage of our function search with an embedding. We tested a narrative-tuned embedding and a general open embedding against the word search on 528 scenes.
What we learned
In two runs both embeddings gave the writer better shapes than the word search.
1 test ended in this step: 1 passed.
2026-10-26 · The core and its workflow · Planning from the library
What we tried
We tried to move our series planning steps into a new tracked pipeline. We also had our writing model plan one full book from the top down and then revise that plan once.
What we learned
The first book plan missed its target because the brief gave only the budget. A full brief and one revision gave a better structure but a review still found gaps in cause and effect. We dropped the pipeline move and no reader has read the plan yet.
1 test ended in this step: 1 dropped.
2026-10-26 · The gate and the reader panel · Judging our work
What we tried
We showed two model judges our first three episodes blind beside episodes of human-written serials that we chose for comparison. Publishing waited until the judges could not tell which pages a machine wrote.
What we learned
Our episodes failed with about 90% machine calls because the judges saw clean and polished prose. The human episodes drew about 60% machine calls only in a test that used their openings at the same story position.
1 test ended in this step: 1 did not pass.
2026-10-26 · Copy edit and continuity · Several theories
What we tried
We planned to render each episode as synthetic speech and flag hard names, long sentences and unclear speaker turns. Only the text checks ran because we had no access to the voice service.
What we learned
Our dialogue lacks a speaker tag in 58 percent of paragraphs against 36 percent in the human-written serials we compared with. We now check this rate on new episodes. We have not yet confirmed it on real audio.
1 test ended in this step: 1 no clear answer.
2026-10-26 · Borrowed pages · Prose that reads human
What we tried
We moved scenes from old novels into the first three episodes of our serial. Two model judges then tried to name the source of each scene in these episodes and in our published episodes.
What we learned
The judges named the source in 7 of 18 scene readings but named no book for our published episodes. Copied sentences keep a source recognisable for obscure books as well as famous ones. We stopped this method for publishing until a recognition test cannot place any source.
1 test ended in this step: 1 did not pass.
2026-10-25 · Trained voices · Prose that reads human
What we tried
We tuned an open model of 31B on period novels with a romance core and some comic and thriller books. We tested its drift, scenes written in short sections, the untuned model and a lower sampling temperature.
What we learned
Tuning beat the untuned model. Model readers still found its scenes more machine-like than the scenes of our writing model. Short sections and comic or thriller books did not fix its drift after good openings and a lower temperature hurt.
3 tests ended in this step: 1 passed · 2 did not pass.
What we kept
2026-10-25 · Scene index and search · Several theories
What we tried
We tested our search for scenes that fix problems in a story plan on our new scene library. We compared it with the old library on 17 problems and the same books.
What we learned
On the new library our search found fixes about as well as on the old one. It needed no problem labels and we will not rebuild them. The judge may partly favour the new summaries because they are about five times longer.
1 test ended in this step: 1 passed.
Drew on
2026-10-24 · Trained voices · Prose that reads human
What we tried
We tuned small and large open base models on scenes by one author or a blend of three authors from before 1931. They continued a page from a plan or rewrote our writing model's scenes.
What we learned
Each step from 1.7B to 8B to 32B held the scene better and read less like a machine. No tuned model passed our bar. Each restyle pass made the scenes of our writing model read more like a machine.
4 tests ended in this step: 4 did not pass.
What we kept
2026-10-24 · Prose and render · Prose that reads human
What we tried
Our writing model wrote nine planned scenes of our own serial from the plan and the story so far. In one arm the prompt ended with a hidden primer of 400 words of human prose.
What we learned
The primer failed. Our writing model ignored it and machine scores stayed near 0.79 with or without it. The model followed the plan and the story well but its scenes still read as machine prose.
1 test ended in this step: 1 did not pass.
2026-10-24 · The lift · Reading a novel
What we tried
On 12 public-domain books we compared our small labelling model and a second model against our main model on every label type. We then set a confidence bar for each type.
What we learned
The second model agrees with our main model more often than the small model on every label type but one. The small model rarely passes its confidence bar and four types get no bar at all. So the second model must give most labels.
1 test ended in this step: 1 passed.
2026-10-24 · Scene index and search · Several theories
What we tried
We searched our library of 64 public-domain books for scenes that carry the planned events of 9 serial scenes. We tried three versions: events as concrete props, events as general shapes and candidates found by the events.
What we learned
Matching a whole planned scene to one library scene failed in all three runs. A blind check by a model showed that our cheap classifier was too loose. True matches were near zero.
3 tests ended in this step: 3 did not pass.
2026-10-23 · Trained voices · Prose that reads human
What we tried
An open base model and two writing models continued the openings of human scenes from old novels. Then we gave the base model a plan as an old chapter heading above the page.
What we learned
The base model failed both times because it drifted off the scene and ignored the plan. The two writing models read more human when they continued a human page. Their scores also moved between runs.
2 tests ended in this step: 2 did not pass.
What we kept
2026-10-23 · Swap and fit · Prose that reads human
What we tried
We tested rules that set each name form after a name swap in six public-domain scenes. A second run matched people by exact name because a loose match had swapped relatives' names.
What we learned
Both runs failed our bar of half the name errors. The errors fell from 21 to 17 and then to 13. The rules fixed spoken full names but missed old words of address, minor people outside our world and formal family address.
2 tests ended in this step: 2 did not pass.
What we kept
2026-10-23 · Prose and render · Prose that reads human
What we tried
Two writing models each continued 12 pages from classic novels in one run. Judges then rated how machine-like each text read.
What we learned
The test failed because the second model read only 0.01 to 0.02 less machine-like than our main writer against the 0.10 we needed. The earlier gap came from noise between runs. We now compare writers only inside one run.
1 test ended in this step: 1 did not pass.
What we kept
2026-10-23 · Borrowed pages · Prose that reads human
What we tried
We built scenes from public-domain novels as sentence collages, as continuations of a human page and as chains of borrowed pages. We also screened sources for recognition and edited pages to our plan.
What we learned
A human page makes our writing model read less like a machine. Lifted plot pieces keep it on our plan. Everything else failed because borrowed pages broke the story, judges named their source books and the text of the model itself did not help.
7 tests ended in this step: 7 did not pass.
What we kept
2026-10-22 · Foundation · Several theories
What we tried
We tested ways to make our labelling model agree more with a careful reader on scene and chapter labels. We tried adding summaries, sending doubtful units to a second model, rewriting unclear definitions and filling three labels by code.
What we learned
Our labelling model held up on new units at a calibrated bar. The second model failed on the hard units below that bar. Summaries only raised confidence, stricter definitions only fed a catch-all type and code-filled labels were worse.
4 tests ended in this step: 3 did not pass · 1 no clear answer.
Drew on
2026-10-22 · Measures and profiles · Several theories
What we tried
We grouped a set of books by how the two lead characters change each other from the start to the end. We hoped to find shared story paths such as rivalry that turns into respect.
What we learned
The test failed. The groups did not depend on chapter order and genre labels mostly predicted them. The groups only separate couples with a growing bond from pairs that rarely change each other.
1 test ended in this step: 1 did not pass.
Drew on
2026-10-22 · The story ledger · Reading a novel
What we tried
We tested two measures of a character's usual behaviour: repeated reactions to events and expected acts over time. Both ran by code on our story ledger for a set of books.
What we learned
The reaction measure failed because a second ledger build found almost none of the same habits. The expectation measure was stable. It shows only a small return to type and no clear growth yet.
2 tests ended in this step: 1 passed · 1 did not pass.
Drew on
2026-10-22 · The lift · Reading a novel
What we tried
We made the story ledger read the one fixed scene cut of each book. We built three books this way and one book twice to measure run noise.
What we learned
Every ledger scene matched the cut. Two runs agreed on 74% of changes against 54% before. Most events fit in one beat but code still could not derive movement or pattern.
1 test ended in this step: 1 passed.
2026-10-20 · Planning · Several theories
What we tried
We built casts of characters for one story premise with life histories from old public-domain memoirs. Then two model judges read blind pairs of casts with and without these histories.
What we learned
The life histories did not make the choices of the characters look more caused by their lives. The cast with histories won only 58% of blind reads and that gap was within the noise of the judges. So we wrote no test episodes.
1 test ended in this step: 1 did not pass.
2026-10-20 · The story ledger · Reading a novel
What we tried
We put a biography from the 19th century through our novel-reading system without changes. We compared the results with two classic novels.
What we learned
The system can read a biography. Its world rules, character traits, life events and reusable story shapes are usable. It fails on reader questions, letters read as narration and the split identity of the subject.
1 test ended in this step: 1 no clear answer.
2026-10-20 · Scene index and search · Several theories
What we tried
We searched 500 scene shapes by their labelled function instead of by the words of a plan issue. For each of 17 issues a second model picked 5 of the top 30 shapes.
What we learned
The function search scored about four times as high as the word search. Most of the gain came from the second model's pick and only one model judged. The search worked on a newer library without notes on what each scene solves.
1 test ended in this step: 1 passed.
2026-10-19 · Trained voices · Prose that reads human
What we tried
We ran a released open paraphraser that is trained to make model prose read as human. We used it on whole pages by our writing model, on edited sentences in human pages and on scenes of our serial.
What we learned
The paraphraser failed on whole pages and on our serial scenes because it lost meaning and read more like a machine. Its edits in human pages beat the edits of our writing model in one side-by-side test. That test used two books only and the paraphraser made more errors.
4 tests ended in this step: 1 passed · 3 did not pass.
Drew on
2026-10-19 · The reader model · Several theories
What we tried
We tested if model judges can rate the order of a reveal in two rewritten versions of a scene. We also compared unearned twists in our serial with two little-known older books.
What we learned
The first test failed because the judges preferred the smaller edit over the better order. Our serial showed no more unearned twists than the old books but had half as many setups. This second comparison was inconclusive.
2 tests ended in this step: 1 did not pass · 1 no clear answer.
Drew on
2026-10-19 · Planning · Several theories
What we tried
We planned to show each invented rule of our serial in an earlier scene before a plot turn needs it. First we tested if the planner could find the rules that lacked groundwork.
What we learned
A model linker met our bar where word overlap failed but we paused the test before we wrote any episode. About half the missing groundwork is rules that the writer invents on the page. No plan can reach these.
1 test ended in this step: 1 dropped.
2026-10-19 · Copy edit and continuity · Several theories
What we tried
We tested a code check for problems that a world rule solves in the chapter where the rule first appears. We also sent the doubtful answers of a rule-based checker on required story events to a second model.
What we learned
Our serial solves problems with a new rule about three times as often as human books do. The code check failed every bar that we set in advance and can only describe. With only the doubtful answers the second model came close to the results of our main model.
2 tests ended in this step: 1 passed · 1 did not pass.
2026-10-19 · Borrowed pages · Prose that reads human
What we tried
We searched old public-domain novels for human scenes that show five of our invented world rules through action. We wanted to place such a scene before the rule solves a problem in our serial.
What we learned
The test failed with no usable scene for any rule. Models recognised all three scenes that we found. So a model must mostly write the rule into the episode itself.
1 test ended in this step: 1 did not pass.
2026-10-18 · The reader model · Several theories
What we tried
Blind model judges tested where a scene should reveal a fact and how a dialogue turn should be grouped. We then checked if the judges simply remember the original texts.
What we learned
Both main tests failed because the judges prefer the least edited version. They do this even on texts they cannot remember. So these judges cannot tell us where to reveal a fact or how to group a dialogue turn.
3 tests ended in this step: 2 did not pass · 1 no clear answer.
Drew on
2026-10-18 · Prose and render · Prose that reads human
What we tried
We tested five register changes in our writer's prompt against our blind model judges. We also evolved the writer and editor prompts against some judges and tested them on the others.
What we learned
Both failed. The judges spot the sentences of our writing model and called every register draft machine-made in 48 of 48 pairs. The evolved prompts fooled only the judges we tuned against and drew more machine calls from the other judges.
2 tests ended in this step: 2 did not pass.
Drew on
2026-10-18 · Foundation · Several theories
What we tried
We tested if a small cheap model gives the same answers as three larger models on two checks. One asks if a chapter shows an event and one asks if a scene shows its required developments.
What we learned
The small model failed the first check and missed one in six events that the larger models called shown. The second check gave no score because our pages and drafts held almost no missing developments.
2 tests ended in this step: 1 did not pass · 1 no clear answer.
2026-10-18 · The gate and the reader panel · Judging our work
What we tried
We built a panel of detectors for machine prose from three model judges, a style test and a small open model. We tested it on human passages, pastiche, our pages and other systems' pages.
What we learned
The style-statistics test found machine prose best and made errors unlike the model judges. Two model judges missed the same pastiche and the third followed one of them. No machine passage escaped the whole panel.
1 test ended in this step: 1 passed.
Drew on
2026-10-17 · The reader model · Several theories
What we tried
We tested if two models recognise famous novels after a name swap and if many model guesses predict the next chapter. We also checked our sentence rhythm measures against eye-tracking data.
What we learned
Both models named every famous book despite the swap. Most rhythm measures failed our checks and our difficulty measure barely predicted reading time. Model guesses showed which questions stay open but not their answers.
4 tests ended in this step: 2 passed · 2 did not pass.
Drew on
2026-10-17 · The story ledger · Reading a novel
What we tried
We extended the story ledger to track what each scene discloses on every layer and how standing changes. We judged long character aims with hindsight and graded each change by size and reach.
What we learned
Hindsight settled 78.5% of long aims against 8% by word matching.
2 tests ended in this step: 2 passed.
Drew on
2026-10-17 · Borrowed pages · Prose that reads human
What we tried
We tested whether our writing model can fit scenes from old human novels to the needs of our own serial. We compared them with scenes that it wrote from story beats only.
What we learned
The swap path gave no usable scene in 13 tries and the control gave 4 in 10. The old scenes were too short, kept old nouns and lacked the needed events. After heavy edits they read as machine text.
1 test ended in this step: 1 did not pass.
2026-10-16 · Measures and profiles · Several theories
What we tried
We added goal spans, reason links and better scene presence to the story records of a set of books. We tested if word matching can tell when a long goal is settled.
What we learned
Word matching settled only one long goal in ten. We must read goals with hindsight instead.
1 test ended in this step: 1 did not pass.
Drew on
2026-10-16 · The story ledger · Reading a novel
What we tried
In three classic novels code linked the needs of each event to the earlier event that met them. We also read characters as state machines and found each book's world rules in one model pass.
What we learned
Links in the real order beat a shuffled order. Character dispositions predicted most chosen acts even on held-out acts. World rules closed most open inputs on state and goals but missed inputs on presence and need a retry.
3 tests ended in this step: 3 passed.
Drew on
2026-10-16 · Copy edit and continuity · Several theories
What we tried
We wrote code that flags a title, colour, owner or source that two passages state in different ways. We tested it on errors that we planted in a set of books and on real errors in our own serial.
What we learned
The code found 3 of 11 real errors in our serial. Together with a check by a model it found 9 of 11. It caught most planted errors and gave about one false flag in 100,000 words.
2 tests ended in this step: 2 passed.
2026-10-15 · Planning · Several theories
What we tried
We tested added planner instructions for book options on the ending, on provisional assumptions and on both. We compared them with the base prompt in 12 runs on one story setup.
What we learned
No reviewed book option was ready to use with or without the added instructions. The added instructions used up to twice the output. We dropped them as defaults.
1 test ended in this step: 1 no clear answer.
2026-10-15 · Measures and profiles · Several theories
What we tried
We retold our own story candidates and a set of published books in a nine-line story form. We compared the results with a model judge of story interest.
What we learned
The form did not agree with the judge. Its failures followed outline length. It penalised the published books as often as our candidates and we dropped it.
2 tests ended in this step: 2 did not pass.
2026-10-15 · The gate and the reader panel · Judging our work
What we tried
We built a test with known answers to check our model judges. It used 90 machine-matched passages from little-read novels before 1931 and 36 versions with planted story faults.
What we learned
Judges that read one passage alone spot machine prose but miss good imitations. Paired questions always gave one-sided answers. Models knew 77% of the little-read books we probed and one model searched the web until we turned this off.
1 test ended in this step: 1 passed.
Drew on
2026-10-14 · Prose and render · Prose that reads human
What we tried
A second writing model wrote the first three episodes from the same input as our usual writer. We also gave the writer a list of habits that blind judges spot in our pages.
What we learned
The second model did not beat our usual writer in the blind test. The habit list gave only a small unsteady gain. No change to the prose step fixes the problem because the cause comes before the prose.
2 tests ended in this step: 1 did not pass · 1 no clear answer.
2026-10-14 · Plan checks · Judging our work
What we tried
We wrote code to check a story plan for cold threads, flat episodes and turns without a setup. Each check had to find planted faults in 80% of cases before we used it.
What we learned
All checks passed the planted tests. Our second arc left threads untouched for four or more episodes at more than twice the rate of the human-written serials we compared with. Flat episodes and turns without a setup did not separate our serial from those serials.
1 test ended in this step: 1 passed.
2026-10-14 · Planning · Several theories
What we tried
We took only the plot beats from human-written books for the first three episodes of our own serial. Our writing model then wrote the full text from those beats.
What we learned
This was the only method that lowered how often blind judges called our serial machine-written. The rate stayed above the rate for the human-written serials we compared with. The judges recognised no source book in 18 scene readings.
1 test ended in this step: 1 passed.
2026-10-14 · The lab's record · Several theories
What we tried
We built one register for all our ideas. Each entry holds the hypothesis, stage, target, experiment, result with interval and status.
What we learned
The register holds 100 ideas and 56 experiments. Its tests pass. Most ideas are still untested: 68 are untested, 29 are touched, 2 are tested and 1 has no assignment.
1 test ended in this step: 1 passed.
2026-10-14 · Scene index and search · Several theories
What we tried
A classifier labelled over 4,000 scenes with the feeling each leaves with its reader. We tested if the labels separate chapter ends from other scenes and agree with a model reader.
What we learned
Chapter ends leave less curiosity and more grief, unease and tenderness than other scenes. Agreement with the model reader met our bar only when we counted dread and unease as one feeling.
1 test ended in this step: 1 passed.
2026-10-14 · The gate and the reader panel · Judging our work
What we tried
We asked a second model to fix typos and small slips in passages from human-written books. Then our model judges compared these cleaned passages with our own episodes.
What we learned
Typos do not explain why the judges call our episodes machine-written. The cleaned passages read a little more human but our episodes scored the same. We dropped the idea.
1 test ended in this step: 1 did not pass.
2026-10-14 · Copy edit and continuity · Several theories
What we tried
We put old continuity slips back into our own serial and tested two continuity checks on them. We also tested a copy edit by code only.
What we learned
Strict rules missed our continuity slips because these slips are wrong details such as a title or a look. A direct read by a model found more slips. The code copy edit found every planted error.
3 tests ended in this step: 1 passed · 2 did not pass.