We Run 9 Checks on Every AI Chapter Before You See It. Here's What They Catch.

Every AI writing tool generates text. That is the easy part. The hard part -- the part almost nobody does -- is checking whether the text is actually consistent with your story.

At Novarrium, every drafted chapter passes through 9 automated checks before the author sees it. What happens when one fails depends on what failed. A contradiction of a fact you wrote or confirmed is repaired in place or the chapter is redrafted, and if it is still there the chapter is held for you. A judgment call, or a fact the system inferred and you never confirmed, comes back as a note that quotes the line. We would rather hold a chapter than hand you a broken one, and we would rather ask you than guess.

This is not a marketing claim. It is the core architecture of the product, as it runs today. Here is exactly what each check does.

Check 1: Character Traits Locked

Every character in your story has a profile: physical appearance, personality traits, backstory, relationships. These details are stored in the Story Bible as structured facts, and you can lock the ones that must never change.

After drafting, this check compares character descriptions in the new chapter with the established facts. Colors (eyes, hair), build, and left or right (a scar on the left hand) are checked by code, not by a model. Other traits are read by a verifier that has to quote the line it objects to. If your protagonist has green eyes in chapter 1 and plain narration gives them brown eyes in chapter 12, that is a contradiction: it is repaired, and the chapter is held if it cannot be.

Prose often says how a trait looks, not what it is: "in the firelight his eyes burned gold". A line framed by light, comparison, memory or speech comes back as a note, never a silent rewrite. Ages are notes too, because a flashback or "looked forty" is not a contradiction.

Check 2: Dead Characters Stay Dead

When a character dies, the Story Bible records it with the chapter it happened in. Before a chapter is drafted, a check warns you if the plan puts a dead character in a scene. After drafting, the chapter is checked for dead characters acting in the present.

The character can be referenced in past tense -- memories, flashbacks, other characters discussing them. But they should not take actions, speak dialogue, or be physically present in scenes. If a chapter brings back a character whose death you wrote or confirmed, it is redrafted or held. If the death is one the system inferred and you never confirmed, you get a question instead: which is right?

This is one of the most common failures in AI writing, and one of the most destructive to reader trust. A character death that gets undone without explanation demolishes the emotional arc of the story.

Check 3: POV Perspective Checked

If your chapter outline specifies first-person POV for a chapter, the writer is told so, and the drafted prose is checked against it.

AI models default to third-person limited because that is the most common perspective in their training data. The check looks for a chapter that does not match its plan: the wrong perspective, or a chapter that never names its POV character and centers someone else. A mismatch comes back as a note for you to review rather than an automatic rewrite, because multi-POV books break the simple rules on purpose.

Multi-POV novels work the same way: chapter 1 might be first-person Isolde, chapter 2 third-person Leila. Each chapter is checked against its own plan.

Check 4: Plot Points Verified

Your chapter outline contains key events that must happen: confrontations, revelations, decisions, arrivals. They go into the prompt, and after drafting this check records which ones the chapter contains.

If your outline says "Isolde discovers the letter" and the drafted chapter never mentions the letter, you see it on the chapter. A missed beat is a judgment call (a beat can happen off the page, or in other words), so it is a note for you, never a block.

This catches a common failure mode where the AI drafts pleasant prose that reads well but ignores the plot beats the author specified.

Check 5: World Rules Scanned

World-building rules are constraints on what can exist in your story's universe. A humans-only world should not have elves. A world without gunpowder should not have firearms. A magic system with specific rules should not be violated.

Only your rules are checked: Novarrium has no genre defaults, so a phone in your urban fantasy is fine unless you wrote that it is not. A code scan looks for the words your rules forbid. A hit inside a comparison ("like a radio"), a negation ("without a phone") or a wish goes to a closer read instead, and rules the scan cannot read are judged by a verifier that must quote the line. A plain violation of your rule is repaired or holds the chapter; an uncertain one comes back as a note.

Check 6: Voice Consistency Matched

Every character in your novel should have a distinct voice. A grizzled military commander does not speak like a teenage apprentice. A formal diplomat does not use street slang.

Each character's personality profile, built with Big Five (OCEAN) scores, and their speech patterns go into the prompt. After drafting, a voice check reads the chapter against your style settings and the characters' established voices. Voice is a judgment call, so a drift is a note for you, not a hold.

Check 7: Author Rules Obeyed

Authors can set custom rules for their project: "no romance subplot between Isolde and Orla," "never describe violence graphically," "maintain a hopeful tone throughout." These are given to the writer and checked after every chapter.

Author rules are checked by a targeted read that evaluates the prose against each rule and must quote the line it objects to. This is not keyword matching -- it is a contextual read of whether the rule was followed.

A rule about what exists in your world (a world rule) holds the chapter when it is broken and cannot be repaired. A rule about style or tone ships with a note quoting the line, because whether a scene is "graphic" or "hopeful" is your call to make.

Check 8: Quality Scored

Every drafted chapter receives a quality score from 0 to 100. The score evaluates prose quality, pacing, dialogue naturalness, and narrative coherence.

A chapter below 75 gets a revision pass with the score's feedback; one below 50 gets another attempt. A chapter that is still below the line is not held for its score alone: the score is shown to you with the chapter.

This keeps the engine from settling for technically consistent but poorly written prose. Consistency without quality is not useful. The chapter needs to read well and be consistent.

Check 9: Story Bible Updated

After a chapter is drafted, the Story Bible is updated with the new facts it introduced. New characters, new locations, new relationships, new events -- extracted and stored for future chapters. The same happens when you save edits of your own.

This is what makes the pipeline cumulative. Chapter 1 establishes facts. Chapter 2 is checked against those facts and adds new ones. A fact the system extracts is marked as the system's: it never overwrites one you wrote, and it is never locked until you confirm it. That is deliberate. The system can misread your prose, and a misread fact should raise a question, not block your book.

The extraction is automatic. The author does not need to maintain a character sheet or a spreadsheet; reviewing the facts the system found, and locking the ones that matter, is the part worth your time.

The Retry Loop

When a check finds a contradiction, the first move is a repair, not a rewrite. A single wrong token (an eye color) is swapped by code. A single sentence is patched, re-checked and kept only if the check now passes. Only when a repair cannot fix it is the chapter redrafted, with specific feedback about what failed: "dead character Malik appeared in dialogue," "the rule 'no firearms' is broken by 'she drew her pistol'."

This feedback loop is critical. The AI does not just try again blindly. It knows exactly what went wrong. A chapter gets up to two redrafts.

After that, a chapter that still contradicts a fact you wrote or confirmed, or a world rule you set, is held. We do not quietly ship it. You are told what it contradicts and decide: fix the chapter, change the fact, or redraft with different direction. Everything else ships with its notes, so you know exactly which lines to read.

Why Most Tools Do Not Do This

Verification is expensive. Running 9 checks adds processing time and compute cost to every chapter. Most AI writing tools optimize for speed: generate fast, deliver immediately, let the author catch the errors.

That approach works for short content. Blog posts, emails, marketing copy -- generate and ship. But novel-length fiction has a unique problem: errors compound. A character trait that drifts in chapter 5 creates cascading contradictions through chapters 6 through 20. A missed plot point in chapter 3 derails the entire storyline.

The cost of not verifying is far higher than the cost of verifying. A single undetected contradiction can require rewriting multiple chapters. A pipeline that catches it before the author reads the chapter saves hours of revision work.

The Bottom Line

AI writing tools that generate without verifying are doing half the job. The generation is the easy part. The verification -- checking character traits, dead characters, point of view and plot points against your own story -- is what separates a tool from a workflow. It is a net, not a guarantee: read your chapters, and settle the notes.

Nine checks. Repairs before redrafts. Holds for the contradictions of facts you confirmed, notes for the judgment calls. That is how every chapter on Novarrium is built.

Try it free -- see the checks on your own chapters. Your first 3 chapters are drafted free, no credit card required.