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 generated chapter passes through 9 automated verification checks before the author sees it. If any check fails, the AI rewrites the chapter with specific feedback about what went wrong. It gets up to 3 attempts. If it still cannot produce a compliant chapter, we block it from shipping rather than delivering broken prose.

This is not a marketing claim. It is the core architecture of the product. 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.

After generation, this check verifies that character descriptions in the new chapter match the established profile. If your protagonist has green eyes in chapter 1, they cannot have brown eyes in chapter 12. If a character has a scar across their left cheek, it does not migrate to their right cheek.

The check covers: eye color, hair color, height, distinguishing marks, clothing patterns, speech patterns, and any other physical or behavioral trait stored in the character profile. Immutable traits -- things that cannot change -- are enforced as hard constraints. A mismatch triggers a rewrite.

Check 2: Dead Characters Stay Dead

When a character dies, that death is recorded as an immutable fact in the Story Bible. This check scans every subsequent chapter for active appearances of dead characters.

The character can be referenced in past tense -- memories, flashbacks, other characters discussing them. But they cannot take actions, speak dialogue, or be physically present in scenes. If the AI brings back a dead character, the chapter is rewritten.

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 Enforced

If your outline specifies first-person POV for a chapter, the generated prose must be in first person. Not mostly first person. Entirely first person.

This check detects pronoun shifts, omniscient intrusions, and camera-perspective descriptions that break the specified narrative perspective. AI models default to third-person limited because that is the most common perspective in their training data. Without enforcement, POV drift is nearly inevitable in longer chapters.

The check also handles multi-POV novels where different chapters use different perspectives. Chapter 1 might be first-person Kael, chapter 2 might be third-person Maya. Each chapter is verified against its own specification.

Check 4: Plot Points Verified

Your chapter outline contains key events that must happen: confrontations, revelations, decisions, arrivals. This check verifies that every key event from the outline actually appears in the generated prose.

If your outline says "Kael discovers the letter" and the generated chapter never mentions the letter, it fails this check and gets rewritten. The AI receives specific feedback about which plot points were missed and generates a new version that includes them.

This prevents a common failure mode where the AI generates pleasant prose that reads well but completely 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.

This check uses deterministic pattern scanning to detect violations of your world rules. It is not just checking whether the AI mentioned something it should not have. It is scanning for categories of violations: forbidden species, anachronistic technology, impossible geography, and magic system violations.

World rules are among the hardest constraints for AI to maintain because they involve nested dependencies. A magic system where "only women can channel" is a simple rule. A magic system where "only women can channel, except men who were born during an eclipse, but their channeling corrupts them over time" is a web of interconnected constraints. The check handles both.

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.

This check analyzes dialogue and internal monologue against each character's established voice profile, which is built using Big Five (OCEAN) personality modeling. It detects when a character's speech patterns, vocabulary level, or emotional tone drift away from their established baseline.

It also catches invention violations -- when the AI creates new characters or locations that do not exist in the Story Bible. If a character named "Theron" suddenly appears in chapter 7 and was never established, that is flagged.

Check 7: Author Rules Obeyed

Authors can set custom rules for their project: "no romance subplot between Kael and Sera," "never describe violence graphically," "maintain a hopeful tone throughout." These are checked after every generation.

Author rules are verified using targeted analysis that evaluates the generated prose against each rule. This is not keyword matching -- it is contextual evaluation of whether the spirit of the rule was followed.

Author rules have hard-gate enforcement. If the AI violates an author rule after 3 attempts, the chapter is blocked entirely. The author's explicit instructions are treated as law.

Check 8: Quality Scored

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

Chapters below 50 are hard-rejected and rewritten immediately. Chapters between 50 and 75 get one additional generation attempt to improve. Chapters above 75 pass this gate.

This prevents the AI from shipping 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 passes all checks, the Story Bible is automatically updated with any new facts introduced in the chapter. New characters, new locations, new relationships, new events -- all extracted and stored for future enforcement.

This is what makes the enforcement pipeline cumulative. Chapter 1 establishes facts. Chapter 2 is verified against those facts and adds new ones. By chapter 20, the Story Bible contains hundreds of facts, and every single one is enforced on every subsequent generation.

The extraction is automatic. The author does not need to manually update a character sheet or maintain a spreadsheet. The system tracks everything.

The Retry Loop

When any check fails, the chapter is not simply flagged with a warning. It is regenerated. The AI receives specific feedback about what failed: "POV drifted to third person in paragraphs 4 and 7," "dead character Malik appeared in dialogue on line 12," "key event 'letter discovery' was not included."

This feedback loop is critical. The AI does not just try again blindly. It knows exactly what went wrong and generates a targeted fix. Most failures are resolved on the second attempt. Some require a third.

After 3 failed attempts on a hard-gate check, the chapter is blocked. We do not ship broken prose. The author is notified and can adjust their outline or regenerate with different parameters.

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. An enforcement pipeline that catches it before the author sees it 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 -- making sure every character trait matches, every dead character stays dead, every POV stays locked, every plot point lands -- is what separates a tool from a workflow.

Nine checks. Up to three retries. Hard gates that block broken prose. That is how every chapter on Novarrium is built.

Try it free -- write your first chapter and see the enforcement pipeline in action. Your first chapter is free, no account needed.