In A Storm of Swords, Catelyn Stark looks at her son's new wife and judges her hips good for bearing children. In the next book, A Feast for Crows, Jaime Lannister meets the same girl and notes narrow hips. Same character, same attribute, two point-of-view chapters, and a contradiction that nobody between the manuscript and the shelf caught before readers did.
What happened next is the interesting part. Readers did not file it as an error. They built a theory: the girl Jaime met was not Jeyne Westerling at all but a substitute, swapped in by her mother, and the hips were the clue. According to the fan-maintained archive of Martin's public remarks, he told a reader at a 2011 signing that it was a mistake, and the line is reportedly gone from current editions. The theory is still up.
I keep coming back to this one because it is the whole problem in miniature. The error was not in the prose: both sentences are fine. It was not a memory failure in any ordinary sense: the author knew his character. It crossed a book boundary and two narrators, which is the gap no re-read covers. And the readers' first instinct was to repair the contradiction by inventing plot, which is exactly what a writer's own brain does on a re-read, silently.
Here are four habits that catch this class of error, each followed by what it turned into when we built it into Guardian. The second half of each is not a sales pitch. Building a habit as a machine is the fastest way to find out what the habit actually is, and in two of the four the machine is more honest about its limits than the advice ever was.
1. Record facts with their mutability, not biographies
The standard advice is to keep a character bible. The standard failure is that the bible is a biography: a page of prose about who Jeyne is, which you stop updating in week three because updating it is a second writing job.
The version that survives is not prose. It is a list of keyed facts, and the thing you record about each fact is not only its value but whether it is allowed to change:
- Immutable. Birthplace, parentage, species, the city a named institution sits in, the width of someone's hips. A later contradiction is an error, full stop.
- Caused. Residence, rank, employer, marital status, what someone is carrying, a wound. It may change, but the manuscript has to show what changed it. A contradiction with no on-page cause is a finding.
- Free. Mood, clothing, hair colour, what someone had for breakfast. Contradiction is normal life and should never be reported.
Three classes, because three is how many different things you can do with a contradiction: flag it, go looking for a cause, or ignore it.
The reason to bother is what happens without the axis. A checker that does not know which facts may change has two settings, flood or retreat. Our first attribute rule chose retreat. It compared only physical traits, which is how a university that quietly moved provinces between chapters went unnoticed while a hair colour change did not.
What it became. Every recorded fact in Creader carries a mutability class, and the contradiction detector reads it before deciding whether to speak. Rows written before the column existed are read as caused, the conservative middle: reading them as immutable would have turned every legitimate change in every existing book into an error on the day it shipped, and reading them as free would have exempted them from checking forever.
There is a second ledger, and it answers a different question. The world log is not "what is true about Jeyne" but "what could a later chapter reverse". Its extractor gets a budget of six facts per chapter and is told to spend it in priority order: where a person was and when, whether someone is alive or reported dead, the state of objects and doors, who someone is, what someone knows or has claimed. A duration, a routine, how loud a clock is: texture, almost never reversed, not worth a slot. The prompt says in so many words that a tidy but shallow ledger is a failure. Jeyne's hips would not make that ledger. They belong to the first one, under immutable.
2. Keep a delta log, one chapter at a time
Three minutes after you finish a chapter, write down only what this chapter changed. Not a summary. Not what happened. What is now true that was not true before, what stopped being true, and, the column writers skip, who said so.
That column is the one that would have flagged Jeyne. Catelyn's chapter does not establish a fact about hips. It establishes that Catelyn judged them. Jaime's chapter establishes that Jaime saw them. Same key, two claimants. The moment you write the second line you are looking at the first, and the contradiction is on the page in front of you instead of in a forum post two years later.
What it became. The world log is this delta log, generated by a model and folded by code. Each chapter appends one event. An event carries assertions, each keyed by subject and attribute with a by column for the claimant, so that two witnesses who disagree land on the same key and the disagreement is visible rather than filed as two unrelated facts. It also carries refutations, in three kinds a hand-kept log never bothers to separate: the world moved on (the ring was silver, now it is gold, and the earlier record was true when it was written), the reader was being misled (the earlier record is overturned and everything that rested on it is tainted with it), or the draft was revised (the record is simply erased). The current state of the world is never stored. It is recomputed from the log on every read, because a stored copy drifts and nothing notices when it does.
The detector that reads it, Canon changed here, is deliberately timid. It makes no model call, finds the records this chapter overturned, and asks a question with a question mark in it. It never says "error". It cannot: the log can tell that a value flipped, it cannot tell plot from slip. Only the writer can, in one click, and each click is a label the next version of the pipeline will learn from. It also stays quiet on one family of keys, where characters are, because people move and a note on every move is the noise this detector exists to measure rather than produce. A reported death overturned, on the other hand, is worth one question.
What I owe you about this one: today Creader builds the log in the background after a deep run, and you do not see it as a list. You see a question when a key flips. The three-minute habit is still yours to keep. The machine's log is the check on it, not the replacement for it.
3. Build the who-knows-what table, and read the swap theory with it
The swap theory is worth taking seriously, not because it is right but because of what it needed in order to exist.
For "the girl Jaime met was a substitute" to be plot rather than error, somebody on the page has to know it. Her mother would have to have arranged it. The stand-in would have to be carrying it. Somebody, somewhere, would have to act on the secret or nearly give it away. Readers went looking for those scenes and found hints they could argue about, which is how fan theories work. But write the table down and it is empty:
| Character | Information | Learned (chapter, how) | Displayed (chapter) | |---|---|---|---| | Sybell Spicer | the girl at Riverrun is not Jeyne | none | none | | the girl at Riverrun | she is standing in for Jeyne | none | none |
No chapter where anyone learned the secret. No chapter where anyone acted on it. That is the test. A twist is a contradiction with a knowledge row. A slip is a contradiction without one.
The table is two columns of chapter numbers. "Learned" means the prose says they were told, or they were in the room when it was revealed. "Displayed" means they acted on it, referred to it, or reacted to it. The failure it catches is the one invisible while drafting, because you know the secret: a character acting on information nobody gave them, or being shocked by a revelation they received four chapters ago. It is the knowledge ledger pass, and the one that finds what beta readers describe as "something felt off".
It is also the one most often confused with point-of-view leakage, which is a different error: a limited narrator reporting what the viewpoint character cannot perceive. That is a sentence-level problem. This is a book-level one, and the two chapters it involves are usually far apart.
What it became. Creader has a Knowledge state panel that scans a chapter and proposes rows of exactly that shape: learned, marked stated or present, and displayed, each anchored to a chapter and carrying a verbatim excerpt, because a knowledge claim with no quote is a fabrication and the extractor drops those before they land. Rows land pending. Nothing reads a pending row. Once you accept one, a rule with no model call does arithmetic over reading order: displayed before learned is a finding; displayed with no chapter anywhere that teaches it is the hardest finding, and it outranks the soft one, learned twice, when the cap of five notes per chapter cuts. The card names the other chapter and quotes it, because sending you off to find a chapter forty pages back is the thing that makes a continuity flag feel like homework.
Two decisions in it are worth stealing for a hand-kept table. Presence comes from reading the scene, not from a cast list: a chapter with three scenes and one revelation would otherwise hand the secret to everyone in scenes one and three, which manufactures knowledge, and manufactured knowledge both hides real findings and invents false ones. And the extractor is given every statement the book already records, oldest first, and told to copy a matching one back character for character rather than rephrase it. Oldest first, because the secrets at risk are the long-range ones, planted in chapter 2 and paid off in chapter 30, and a newest-first window pushes exactly those out of view.
Belief is a different ledger again: not what a character knows but what they hold, including what they hold wrongly. We built that extractor too, and most of it is still switched off.
4. On the re-read, read facts, not prose
The last habit is about what you look at when you go back. Most writers re-read by chapter, front to back, and are reassured, because every chapter is internally fine. Jeyne's two descriptions are both good sentences. Nothing in either chapter is wrong. The error exists only when the two are placed next to each other, on one key.
So re-read by entity, not by chapter. Pull every line about Jeyne, in order, with its chapter number, and read the list. Then every line about the sword, the wound, the locked door. The prose will try to distract you; it is prose, that is its job. Read the values.
What it became. This is the design rule behind every continuity note Guardian shows. A note has two halves rendered as siblings: the flagged passage, from the chapter you are in, and You established this in …, the earlier chapter's own words. The judgement a writer has to make is a comparison, so the note shows both sides of it. The detector that reads the accepted fact ledger against the chapter is a model call on a deliberate deep run; the attribute rule that runs while you type makes no model call and reads only physical traits, which is the gap the mutability axis was built to close. In both cases the verdict is yours. Nothing is rewritten. Accepting a fact records it against the chapter that established it and leaves the character's card alone, because a card cannot say "true from chapter twelve on", and that qualifier is the entire point.
Why habits and machines, both
Each of the four is a habit that works on paper and that most writers drop by chapter forty, because the cost is paid every chapter and the benefit shows up once, at chapter 140, when a reader with a good memory posts a theory.
Building them taught us where each habit's real content was. The first is not "record facts"; it is "classify how each fact may change". The second is not "summarise the chapter"; it is "write down who claimed what". The third is a table with two columns of chapter numbers. The fourth is a reading order. A machine can keep the first, second and fourth running while you type, and can hold the third once you have confirmed its rows. What it cannot do is decide whether a flipped value is a twist or a slip. That click is yours, and it is the most valuable thing you do with any of this.
Jeyne's hips were a slip. The theory was better than the error deserved. The point of all four habits is to make sure the next one is caught by you, with a chapter number, rather than by someone else, with a theory.