What AI consistency checking can and cannot do

A consistency checker sounds like a machine that reads your book and tells you what is wrong with it. The reality is narrower, and more useful once you accept it: the checker compares what a chapter says against a recorded body of facts, and flags the disagreements. Inside that lane it can do work no human reader would sustain across three hundred chapters. Outside that lane it has nothing to say, and a tool that pretends otherwise is selling you a number, not an evaluation.

What a checker can actually catch

Everything in this list is mechanically checkable, because each item is a comparison between new text and a record that already exists:

• Character contradictions. Chapter seven establishes that Kai cannot use fire magic; chapter eighteen has him casting it with no exception recorded. State changes are among the most checkable facts in fiction (see tracking character state).
• Timeline clashes. Two scenes place the same character in two cities on the same date, or a flashback collides with the present (see timelines across POVs and flashbacks).
• Object continuity. A destroyed sword drawn again nine chapters later.
• Knowledge errors. A character acts on information that no previous scene established they have (see tracking what characters know).
• World-rule violations. The story established that humans cannot teleport; someone teleports (see Magic A is Magic A for why readers hold this against you).

These are exactly the errors long fiction dies of, and exactly the errors a human author stops catching reliably somewhere around chapter thirty.

The record is the ceiling

A checker can only check against what has been recorded. If a rule was never written down, then as far as the checker knows, no violation exists. The tool's power is capped by the quality of the story bible feeding it, which is why tools that treat the bible as a first-class, growing record beat tools that bolt a "check my story" button onto raw text. Raw text has no canon to check against: two vague sentences in different chapters are not, to any machine, a contradiction.

What no checker can do

• Taste. Whether a sentence is good, whether a joke lands, whether the prose has voice. There is no measurement of "good scene" that survives contact with a second reader, and a tool that grades your chapter 87 has generated a number, not an evaluation. Kanonara's design stance here is explicit: severity levels for contradictions with evidence, and no fake objective story-quality scores.
• Pacing and structure. Whether the middle sags, whether a reveal arrives too early. These judgments need a reader who can be bored.
• Intent. Whether a contradiction is a mistake or a mystery. An intentional reveal looks identical to an error from the outside: both are new text that disagrees with the record. This is the false-positive problem, and it is not a bug awaiting a fix; it is a permanent fact.

False positives are the system working

A planted contradiction (a character lying, an unreliable narrator, a mystery the book means to open) triggers the same flag as a genuine mistake. That is unavoidable, and it is why severities exist: a direct contradiction with strong evidence is critical, something merely unclear is a question. It is also why the resolution actions belong to the author: review the evidence, mark it intentional, ignore it, correct the story, update the record, or write the explanation that dissolves the contradiction. The machine proposes the question; only you know the answer.

Older media have known this division for a century. Film and television employ a script supervisor whose whole job is continuity, tracking props, costumes and the timeline against documentation (sometimes literally a story bible), precisely because scenes are not shot in order and human memory fails. The checker's role in prose is the same: tireless comparison, no judgment. The judgment was never delegable, and the oldest long works already show what happens without it; readers have been spotting continuity slips in Homer for over two thousand years.

Working with one honestly

• Treat every flag as a question, not a verdict. Read the evidence it points to before deciding anything.
• Record your resolutions. "Marked intentional" with a note is canon maintenance; ignore-without-reading is how real errors hide in the noise.
• Feed the record. Every scene you write establishes facts. Get them into the bible in the same session you wrote them, or the checker is guarding half the story.
• Keep the meaning side human. Even the platform most relaxed about AI content requires that AI-generated work "be moderated and refined by the author to ensure quality, continuity, and readability." The checking can be machine work. The caring cannot.

A consistency checker will not write your book, will not tell you whether it is any good, and will not notice that your villain became boring in chapter twelve. What it will do is remember every promise your text made and hold each new chapter to them. That is a smaller promise than the marketing suggests, and a bigger help than the skeptics expect.