Checking a Novel for Contradictions: Whole-Book Consistency Check

In this article 7 sections

A highly upvoted comment can drive an author mad: “Chapter 12 clearly says the protagonist’s left hand is in a cast, so how is he swinging a sword with both hands in chapter 47?” You flip back and look, and it’s true. An injury casually planted over a hundred chapters ago, long since forgotten.

This kind of thing is far too common in long novels. A character was at the south-city inn last week and suddenly appears in the north-city tavern this week, with not one line of transition. A magic system agreed three volumes ago that fire can’t be used, and in volume five, carried away by excitement, a supporting character conjures a fireball. Readers have sharp eyes: continuity errors, broken threads and mismatches get screenshotted and posted in the comments, and a dozen-plus of them can make a hot forum thread.

Flipping through the manuscript by hand can’t keep up. What to do? Maliang Writer’s recently launched “Whole-Book Consistency Check” targets exactly this problem: one click scans the whole book, pulls out the contradictions and lines them up as a list, each with an excerpt from the original text and a chapter location, and one click jumps back to the editor to fix it.

Why Is Self-Checking a Million Characters by Hand Unrealistic?

Do the time math and you’ll see. For an 800,000-character book of two hundred-odd chapters, you’d have to check chapter by chapter against character settings, worldbuilding rules, foreshadowing clues and timeline nodes. Just pulling that information out of your head and lining it up is not something a person can do. Late in the book you can’t even remember whether the supporting character in chapter three is called “Lu Heng” (with the character meaning “balance”) or “Lu Heng” (with the character for a fragrant herb), let alone work out which foreshadowing has been paid off and which is still hanging.

Readers can find contradictions because they only watch one spot. An author’s self-check must watch all the cross-relationships in the whole book, and the workload grows geometrically. This job shouldn’t be done by a person; it should be handed to a system.

Some authors set up an Excel sheet to record character settings and foreshadowing status themselves. At first they can keep it up, but after fifty chapters the sheet goes to ruin: every chapter requires updating a dozen-plus rows by hand, who can stand that? And the sheet can only record information you actively enter; details that pop up when you’re writing excitedly don’t get recorded in time, and it is exactly these omissions that become the minefield of later continuity errors.

What Does Each of the Four Dimensions Scan?

The check covers four dimensions, and after each scan the contradiction list is output grouped by dimension.

The character and ability dimension watches whether characters’ basic settings contradict each other. For example, chapter 12 says the protagonist’s left hand is injured and in chapter 47 he wields a sword with both hands: a typical character continuity error. Or a supporting character who appears in chapter two as a “fire-element mage” suddenly unleashes a water-element finishing move in chapter sixty, with no class change or training explained in between. Readers catch this kind of glaring flaw every time, but late in the book the author can’t remember what they planted.

The worldbuilding dimension checks consistency at the rule level. In volume three you set “spiritual power below tier ten cannot fly”, and in volume eight you let a tier-five character fly all over the sky; the check flags this contradiction. Or the settings say “currency is in spirit stones”, and later “gold coins” suddenly appear as the price unit with no transition explained. Rule self-consistency is the lifeline of a long novel; break one place and a whole stretch collapses.

The foreshadowing payoff dimension is my favorite block in this feature. The system automatically identifies the whole book’s foreshadowing, then tells you which has been paid off and which is still hanging unresolved. Chapter 18 plants “the birthmark on the protagonist’s back seems to be heating up”, and it is never mentioned again through to the ending: with this kind of broken foreshadowing, readers are still waiting for the follow-up at the last chapter, and you forgot. The check list has a dedicated foreshadowing zone, and unpaid foreshadowing can be registered into foreshadowing management in one click, so you can focus on tying them up later.

The timeline dimension is marked “Experimental”, and currently checks whether the order and span of events are reasonable. For example, the protagonist goes from entry level to the Golden Core stage within three months while the earlier setting says Golden Core takes at least three years; or an event happens in winter, and the very next chapter’s plot describes “cicadas buzzing”. When seasons don’t match, the timeline dimension raises a flag. This dimension is still iterating and occasionally gives false positives, but it won’t miss major timeline flaws.

Each issue comes with three things: an excerpt from the original text (a direct snippet of the sentences where the contradiction occurs), a chapter location (one click jumps to the corresponding spot in the editor), and a one-line fix suggestion. You don’t need to flip back and forth through the book for context; the list itself is a to-do work order.

One thing to state specifically: among the results of the four dimensions, character and worldbuilding are hard-rule comparisons with the highest accuracy; foreshadowing payoff relies on semantic understanding and occasionally has blurry borderline cases (for example, whether a certain passage counts as foreshadowing), but it won’t miss the big picture; the timeline dimension is marked “Experimental”, meaning it is still being optimized, so accept that it occasionally acts silly.

How Is the Consistency Score Calculated? How Do I Use the Report?

The report gives the whole book a consistency score out of 100. The deduction rules are simple: 8 points per hard error, 3 points per suspected contradiction, 1 point per note-level item. If a two-hundred-chapter book turns up 5 hard errors and 10 suspected ones, the total is around twenty or thirty, and frankly, you won’t feel at ease before handing that score over.

The list is arranged in three severity levels: hard errors, suspected, notes. A hard error is an iron-clad self-contradiction, which must be fixed; a suspected one is where the system thinks there is probably a problem but needs your manual confirmation, such as two names being very similar or a setting described with subtle differences before and after; a note is more like a friendly reminder, such as a certain foreshadowing that has gone eighty chapters without payoff, urging you to take notice.

On the report page, each issue can be marked “Handled” or “False positive”. Items marked handled collapse out of the list, and false positives leave a record so they aren’t reported again next time. A hard error you can’t fix? Send it to the AI assistant in one click and let it draft a revision plan based on the context.

There is also a very practical feature called historical report comparison: the system keeps your last 10 check reports, and each can be compared with the earlier ones, telling you which problems are newly appeared, which have been fixed and which old problems remain. Looking back to check once midway through a serial lets you see intuitively whether the book’s consistency is getting better or worse.

The foreshadowing zone deserves a few more words of its own. Foreshadowing management is the biggest headache in a long serial: when planting you just write it casually, and by payoff time it’s long forgotten. The report’s foreshadowing zone lists all unpaid foreshadowing in the whole book together, each marked with its planting chapter. One glance tells you which thread has been broken for how long and which should be fixed first. Think a certain one isn’t foreshadowing and needn’t be paid off? Just mark it away. Foreshadowing you decide to keep can also be registered in the foreshadowing management module in one click, so you can keep tracking it as you write. For more technique on how to plant and pay off foreshadowing, the Foreshadowing and Suspense Guide and Storylines and Plot Nodes have dedicated breakdowns.

Why Run a Check First When Coming Back After a Hiatus?

Authors who return to continue writing after three months, half a year or even longer are in an especially awkward spot: they’ve nearly forgotten their own book. What foreshadowing was planted earlier? What did the settings of each character say? Where has the timeline got to? They have to gnaw through the old manuscript again before they can pick it up.

The “Hiatus Return Briefing” of the foreshadowing payoff dimension is prepared for this scenario. Run one check, and the briefing tells you how many pieces of foreshadowing the whole book contains, how many have been paid off and how many are still hanging. Which foreshadowing was planted before the hiatus and has gone unexplained since: these are the loose threads you must handle first when you resume. Getting the whole picture before you write beats finding out two chapters later that you’ve committed another continuity error.

One author, returning after a half-year hiatus, found only after running the check that in chapter 120 he had planted foreshadowing for a key item and not said a word about it in the following thirty chapters. Without this briefing, he might have kept writing and planted new loose threads, the snowball growing until he couldn’t rein it in.

How Does It Relate to the Knowledge Graph?

Many longtime users know Maliang Writer also has the knowledge graph to prevent continuity errors feature, and ask: with the graph, why the check?

Two things, with different roles. The knowledge graph is a preventive measure during writing: as you write, it weaves character relationships, setting rules and time nodes into a web and reminds you in real time when you write into a contradiction, which is “prevention before the fact”. The whole-book consistency check is a census measure after the draft is finished: once the whole book is written, it cross-compares all chapters at once and catches the cross-chapter contradictions the graph may have missed during writing, which is “filling gaps and plugging leaks”.

An analogy: the graph is the safety officer on the construction site, and the check is the acceptance inspection after completion. Using both is what counts as double insurance. To understand the approach to long-novel consistency maintenance more systematically, the Long-Novel Consistency Guide goes into more detail.

How Do Pricing and the Free Trial Check Work?

The free part first: the first 3 chapters of each book can be checked once for free, 0 credits. When starting a new book, run the free quota first to get a rough sense.

The formal whole-book check is billed by character count, 20 credits per 10,000 characters. Anything under 10,000 characters starts at 10,000, i.e. a minimum of 20 credits; for a new book we still suggest first using the free check of the first 3 chapters. For 800,000 characters it is about 1,600 credits. The per-run cap is 1,500,000 characters, and longer books need to be submitted in several runs. If a scan fails midway, you are refunded in proportion to the actual completion. The in-site live quote is authoritative.

Subscribers get extra benefits: members at the Pro tier and above get 1 whole-book check per month as a gift, up to 1,000,000 characters per run, which covers most long novels. If you’re wondering whether a subscription is really worth it, Free vs Paid Compared can help you work out the math.

On speed, a 400-chapter book is expected to produce its report in 30 to 45 minutes. The submission dialog shows a dynamic estimated range, not a fixed number. The check is a background task: closing the page after submitting doesn’t affect execution, the Task Center shows the progress “Scanned X/N chapters”, and a notification is sent when the scan is done.

A plain word on the technology: the system first uses an inexpensive model to scan the whole book in batches and pick out candidate contradictions, then uses a long-context model for a final review, pushing false positives down. This two-level mechanism is far more accurate than scanning with a single model. Deeper cross-chapter logic scanning is still in internal testing, and what the first release covers is the four dimensions described above.

FAQ

Q1: What if a false positive turns up?

Every issue in the list can be marked “False positive”, and after that the same item won’t be reported again in the next check. Because the timeline dimension is an experimental feature, it occasionally gives false positives; confirm manually and mark them, which doesn’t affect the accuracy of the other dimensions.

Q2: For books of how many characters is this feature worth it?

The formal check is billed at a minimum of 10,000 characters (20 credits); for a new book you can first use the free check of the first 3 chapters, and decide whether to run the paid whole-book check after confirming the report suits you.

Q3: How long does it take to scan a book?

A 400-chapter book takes about 30 to 45 minutes. The dialog shows a dynamic estimated range when you submit. The check is a background task, and closing the page doesn’t affect execution; the system sends a notification when it finishes, and the Task Center shows live progress.

Q4: Can free users use it?

The first 3 chapters of each book can be checked once for free, 0 credits. The whole-book check needs credits or a subscription benefit. Pro-tier and above subscribers get 1 whole-book check per month as a gift (up to 1,000,000 characters per run).

Q5: After fixing issues, do I need to rescan the whole book immediately?

Usually not. First fix items one by one according to the report and manually review the affected chapters; if you only changed a few local problems, you can wait until the end of a revision stage before doing a whole-book recheck, to avoid spending time and quota again with every single change.