Can AI Keep Levels and Stats Straight? 5 Chapters, 2 Models

In this article 8 sections

We had two models each write 5 chapters from the same English cultivation-farming reference. Gemini 4 Carbon wrote about 19,500 words of English prose with status screens, and Chapter Diagnostics caught 4 number contradictions involving cultivation tier, attributes, quantities, and dates. DeepSeek V4.1 Flash wrote about 14,400 Chinese characters on the English-language page, added no status screen, and the diagnostics found 2 hard contradictions. Once a novel is full of numbers, someone has to keep the books. Chapter Diagnostics can list the numbers and every place they clash, but whether and how to fix them is still the author's call.

Key Takeaways

• A status screen next to the prose doesn't stop the prose from contradicting itself. At the end of the 5 chapters, Gemini 4 Carbon delivered a summary ledger covering the protagonist's level, four attributes, skill levels, and a 25-tael silver stock. Its silver in and out checked out, but Chapter Diagnostics still found 4 contradictions in the prose itself: a recalled cultivation tier, an attribute timing comparison, a potsticker quantity conversion, and a date count-back.
• In this test, the obvious silver math held up. The errors were in time words and fraction words. In both books, the income and spending, price multiplication and division, and snail-nest counts all reconciled in the diagnostics. What broke from the earlier setup were offhand phrases about time and quantity, like "yesterday" and "half a skillet."
• The page language isn't the prose language. Prompted in English on the English interface, DeepSeek V4.1 Flash still wrote a Chinese farming novel, and it didn't add the status screen the continuation prompt asked for.
• No status screen doesn't mean no number ledger. DeepSeek V4.1 Flash wrote no levels or XP, but the prose still contained hard numbers: wages, prices, snail-nest counts, days of injury, and object sizes. The diagnostics found 2 contradictions here too: the timing for restacking the snail nests was reversed, and the white frost layer was introduced twice as a first sighting.
• Chapter Diagnostics lists the books and flags problems. It doesn't rewrite your draft. After all five chapters were written, we ran one read-only diagnostic. The agent pulled out the numbers scattered across chapters, cross-checked them, marked where they clash, and suggested fixes. The original text was left for the author to decide.

How We Tested It

On November 12, 2026, we opened AI Imitation on the Maliang Writer English homepage, picked Royal Road as the source, and chose the hit serial Beware Of Chicken as the style and genre reference. It's a cultivation-themed farming comedy. The original is under copyright, so this article doesn't quote it.

With the reference chosen, we clicked Imitate prose to generate the opening. Then, in the same conversation, we sent two English messages in turn to see whether the model could write through 5 chapters while holding on to every number set up earlier, followed by a read-only check:

Keep writing until the book has five chapters. Keep every level, stat, skill and item number consistent with what earlier chapters established, and show the status screens where it makes sense.

[Skill:Chapter Diagnostics] Check chapters 1–5: list every level, attribute, skill and item number that appears, chapter by chapter, and point out any place where a number contradicts an earlier chapter. Don't change the text.

The two models tested were Gemini 4 Carbon (thinking: medium) and DeepSeek V4.1 Flash. We verified message by message that both the chat model and the prose-writing model were the tested model, and nobody edited the five chapters after they were generated.

Gemini 4 Carbon: Cultivation Farming with a Status Screen

The English novel Gemini 4 Carbon wrote is titled Do Not Disturb the Kiln, with 19,518 words across 5 chapters: Chapter 1 "Fired From the Forge" (3,882 words), Chapter 2 "Itemized Damages" (3,689), Chapter 3 "Hands Off the Damper" (3,923), Chapter 4 "Three Hundred Tiles Before Sundown" (4,023), and Chapter 5 "Tea for an Angry Old Man" (4,001). Writing Chapter 1 took 688 seconds and 236.61 credits. Continuing with Chapters 2 through 5 took 1,427 seconds and 759.99 credits. The Chapter Diagnostics run took 375 seconds and 197.11 credits. The full workflow came to 2,490 seconds and 1,193.71 credits.

Chapter 1 opens with outer-sect apprentice Pei An, burned and bandaged, handing over his duties at the Registry Hall. The opening, verbatim:

Deacon Meng Kui set down his iron brush and stared across the soot-stained desk at the outer apprentice.
Even by the unforgiving standards of the Sundered Peak Forge Sect, the boy looked thoroughly ruined. Coarse linen bandages wrapped both of his forearms from wrist to elbow, smelling faintly of burn salve and tallow. His left eyebrow was singed down to pale stubble, his rough hemp robes bore scorch marks across the chest, and the third-tier Ember Qi circulating through his meridians flickered as weakly as a wet candle wick in a draft. Pitiful, truly, though hardly unusual in the outer halls after a heavy forging week.
What was unusual was the stamped wooden slip resting between them.

After writing all 5 chapters, the model handed back a closing ledger in its reply, listing levels, cultivation techniques, attributes, skills, and supply balances:

• Level & Cultivation: Level 3 -> Level 6 Hearth-Potter (+2 points per level); Ember-Listening Qi (Tier 3 -> Tier 4), Marrow-Tempering Pulse (Tier 1 -> Tier 2)
• Attributes: Body 8 -> 11 | Control 13 -> 17 | Spirit 7 -> 8 | Hearth-Will 15 -> 17 (0 unallocated)
• Skills: Downdraft Firing (Lv. 4 -> 5), Stoneware Wedging (Lv. 4), Cast-Iron Seasoning (Lv. 3 -> 4, 5 -> 6-Layer Lard Cure), Ash-Glaze Mixing (Lv. 2) -> Mineral & Ash Glazing (Lv. 4), Improvised Pan-Parry (Lv. 1)
• End-of-Ch. 5 Stock & Coin: Monthly Quota Cleared (300 tiles + 20 jars delivered, 15 surplus tiles sold), Jade-White Porcelain Cups: 11 (10 Fine, 1 Spirit-Touched), Sage-Green Celadon Test Jar: 1, Washed White Porcelain Clay: 350 lbs, Cold-Jade Pith: 0.5 chunk, Salvaged Tier-5 Sword Steel: 3 lbs, Coin Pouch: 25 Silver Taels (2,500 Coppers)

In the Chapter Diagnostics run, the agent pulled out a number breakdown chapter by chapter. In Chapter 5, for example, Pei An sells one Spirit-Touched Jade-White Porcelain Cup, filled with 2-copper bancha twig tea, to Elder Huo Lie. The cup count drops from 12 to 11, and after taking 20 taels of silver, his coin pouch goes from 5 taels to 25. The diagnostic's excerpt of the remaining materials and silver ledger for Chapter 5:

• Remaining Materials & Wood: Salvaged Tier-5 Sword Steel: 3 lbs (1 lb previously sheared off; exhibits 30% / 3 lbs out of every 10 pig-iron slag fraud from Batch 14 / Lower Furnace 7, P15, P20, P23, P26, P30, P103); Washed White Porcelain Clay: 350 lbs (Pure) (P33, P101); Cold-Jade Pith: 0.5 chunk (P102); 1 8-lb iron splitting maul used by Su Qinghan to split 2 cords of wood (cutting logs into 4 quarters per swing, P75, P77, P81).
• Currency: Huo Lie pays 20 silver taels (4 5-tael silver ingots = 2,000 coppers), raising Coin Pouch from 5 silver taels to 25 Silver Taels (2,500 Coppers) (P65–P66, P68–P69, P72, P84, P104).

The silver and inventory lines were all laid out clearly, yet when the diagnostic checked the prose paragraph by paragraph, it still caught 4 number contradictions:

1. The starting cultivation tier doesn't match across chapters (cross-chapter contradiction). Chapter 1, paragraphs 2 and 37 establish that Pei An starts out at the third tier. In Chapter 5, paragraph 18, Meng Kui, inspecting the kiln on Friday, recalls the scene from two weeks earlier but remembers him as second-tier:

   Ch. 1, P2: "and the third-tier Ember Qi circulating through his meridians flickered as weakly as a wet candle wick in a draft."
   Ch. 1, P37: "Original Pei An only knew two basic sect arts: third-tier Ember-Listening Qi..."
   Ch. 5, P18: "Meng Kui’s breath caught in his throat. Tier-Four Ember-Listening Qi? Two weeks ago, the boy had barely scraped Tier-Two!"

   The diagnostic's suggested fix: Change "barely scraped Tier-Two!" in Ch. 5, P18 to "barely clung to a flickering Tier-Three!".

2. The Body attribute value and the timing of the point gain are out of sync (cross-chapter contradiction). Chapter 2 states that Pei An's base Body is 8 on Thursday evening, rises to Body 9 only late Thursday night after he settles Su Qinghan's bill, and reaches Body 10 only after Zhao Kun's sword breaks in Chapter 3. But Chapter 3, paragraph 29, set at midnight on Thursday, has his current Body at 9, yet the prose says he moves faster than Body 9 did yesterday:

   Ch. 3, P29: "Moving faster than Body 9 had yesterday, I snatched my twelve-pound Spirit-Iron Skillet off the workbench..."

   The diagnostic suggests: Change "faster than Body 9 had yesterday" in Ch. 3, P29 to "faster than Body 8 had this afternoon" (or "faster than my old Body 8").

3. The potsticker count clashes with "half a skillet" (arithmetic contradiction within one chapter). Chapter 4 first says one skillet holds twenty-four potstickers. Pei An fries two full skillets, 48 in all, and Su Qinghan eats eighteen potstickers, which is three-quarters of a skillet. Three paragraphs later, in paragraph 54, the prose calls those 18 "half a skillet":

   Ch. 4, P54: "watching a dignified Seventh-Tier sword cultivator politely inhale half a skillet of pork-and-cabbage potstickers..."

   The diagnostic's fix: Change "half a skillet" in Ch. 4, P54 to "three-quarters of a skillet" (or change eighteen in P51 to twenty-four / "nearly a whole skillet").

4. The day of the week doesn't match the count-back (calendar contradiction). The first three chapters anchor the story's starting point at Thursday dusk. In Chapter 1, paragraph 39, Pei An says he found Old Flatty "Last Tuesday" (2 or 9 days before Thursday). Eight paragraphs later, paragraph 47's narration says "Twelve days ago, footsteps had crunched across the slag heap." Counting back 12 days from Thursday lands on a Saturday, not a Tuesday.

   The diagnostic suggests: Change "Twelve days ago" in Ch. 1, P47 to "Nine days ago" (matching the previous Tuesday).

The diagnostic also added 3 notes that aren't hard errors but could use polish: Chapter 1 never states how many coppers Pei An starts with for the refunded porridge money (he has 38 left after spending in Chapter 2); how Meng Kui's 40 years keeping the roster fits with his 19 years inspecting kilns; and the rhetorical comparison between Su Qinghan's 3 months of meridian damage and "six months of ice-cave seclusion." The five chapters were left untouched, for the author to decide.

DeepSeek V4.1 Flash: A Farming Novel That Came Out in Chinese

Working from the same English homepage and sending the same two English messages, DeepSeek V4.1 Flash instead produced a Chinese cultivation-farming novel titled Don't Mess with My Goose. (Titles and quotes from this novel are translated into English here.) This wasn't a one-off: the same model also wrote in Chinese when imitating Pride and Prejudice and Sherlock Holmes on the English page.

The novel runs 14,387 Chinese characters across 5 chapters: Chapter 1 "Severance Pay Is a Mud Pit" (3,309 characters), Chapter 2 "Three Iron Spikes" (2,918), Chapter 3 "Two Coppers for a Bunch of Seedlings" (3,031), Chapter 4 "The Bamboo Mat at the Bottom of the Pond" (2,828), and Chapter 5 "Someone Wants to Buy This Land" (2,301). Writing Chapter 1 took 604 seconds and 56.32 credits. Continuing with Chapters 2 through 5 took 841 seconds and 84.36 credits. The diagnostic took 69 seconds and 45.56 credits. The full workflow came to 1,514 seconds and 186.24 credits.

In Chapter 1, a plant-protection worker is transported into the body of Shen He, a menial laborer at the Taia Sect's spirit plant garden. After twenty strokes of the rod, he is handed a patch of muddy, rotten land. The opening, translated from Chinese:

Half his face was buried in wet mud, and his shoulders woke up first.
It wasn't pain exactly. It was the kind of ache you get from being scorched by fire and then covered with a wet cloth, a throb that kept time with his heart. He reached out to straighten the seedlings he had crushed sideways, but his hand stopped halfway: this hand was a size smaller than his own, with a layer of old calluses on the knuckles and black dirt packed under every nail.
Then the memories came up, not seeping in bit by bit but dumped over him like a whole bucket.
Taia Sect, the spirit plant garden, menial laborer Shen He. A whole field of third-grade spirit rice, "Green Awn," had withered overnight, and the steward Zhao De had pinned the accounts on him: twenty strokes, then expulsion. The name was exactly the same as his own, which at least saved him the trouble of changing how he answered to it.

As for why the continuation prompt asked for a status screen and the prose has none, the model explained in its reply when it delivered the draft (translated from Chinese):

This book has no levels, stats, skills, or inventory, so there's no status screen to show. The rule set in Chapter 1 is that there's no system. The protagonist's cultivation comes down to one thing: sitting still, he has no qi at all, but once he starts moving, it circulates on its own. His "growth" shows up only as nimbler hands, more work done, and land that yields more. Sect Master Bai's record is likewise kept in real things (the notch in his tail feathers, the defeated opponents under his claws). Turning this into a system novel with a panel would be a major change that overturns the setup already built in the first five chapters. Say the word and I'll switch.

In the Chapter Diagnostics run, the agent first used a script to pull every number out of the five chapters for comparison. It likewise pointed out that "this book has no levels, stat sheets, skill bars, or system items," and checked prices, snail nests, and arithmetic chapter by chapter. An excerpt of the diagnostic report's number ledger (translated from Chinese):

Bottom line first: this book has no levels, stat sheets, skill bars, or system items. That's the setup from Chapter 1, and none of the five chapters contains words like "level / stats / EXP / ding" (I scanned: 0 hits). So what follows are all the hard numbers the book actually uses. The prefix "New" = first appearance in this chapter, "Carried" = carried over from an earlier chapter, and ⚠ = conflicts with earlier text.

Chapter 2

• Prices (fixed in this chapter): a day of casual labor 3 coppers plus a plain meal; lime 2 coppers per dou (a traditional measure); a bunch of seedlings 2 coppers; boundary-stone fee 8 coppers (writing the deed, carving the stone, witnessing)
• Land: boundary stone "a span and a half square," the three characters for "Shen He's property," planted over a foot deep; the deed is one sheet, folded in half
• Snail nests: originally twelve ⚠ → two wrecked by the eel → he recovers ten

Chapter 5

• Prices (carried over): snails 18 coppers for a dou and a half (consistent with 12 coppers per dou, arithmetic correct); lime 4 coppers for two dou (consistent with "2 coppers per dou"); paper 2 coppers a roll
• Land price: forty taels; he earns 3 coppers a day of casual labor, and "a full year of work wouldn't earn him a tael." Working every day of the year comes to about 1,095 coppers, a bit over one tael, so the claim holds
• Snail nests: ten in the previous section, he checked eight, two left (10−8=2 ✓, which ties back to "recovered ten" in Chapter 2)
• Recovery from injury: three days (matches "three days out of the water" in Chapter 4 ✓)

Arithmetic check: 40 taels ÷ 12 coppers per dou → Chapter 5 says "over two thousand dou" ✓; one and a half dou × 12 = 18 coppers ✓; the 8-copper boundary-stone fee is consistent both times ✓.

The visible copper multiplication and division and the snail-nest subtraction all checked out, but on timing and scene details the diagnostic still found 2 hard contradictions:

1. The timing of the snail nests being "wrecked by the eel" doesn't add up (start of Chapter 2). On the same day, the snail nests are still intact in the morning and only get wrecked by the Emberhead eel in the afternoon. But the start of Chapter 2 already describes restacking collapsed snail nests ahead of time, which throws off the "twelve originally, two wrecked, ten left" sequence that Chapter 3 relies on (paragraph numbers in the quotes below follow the diagnostic report's numbering):

   Ch. 2, P1: "The pond had been cleared for three days... restacking stones on the two collapsed snail nests"; yet P69 of the same chapter says "the ones he had stacked stone by stone a few days ago collapsed in two places."

   The diagnostic's suggested fix: "Change the start of Chapter 2 to 'raking the rotten leaves out of the pond' (delete 'restacking stones on the two collapsed snail nests'), keep the wrecking in P69, and the twelve → two → ten in Chapter 3 all works out."

2. "That very thin layer of white on the wet mud" appears twice, and conflicts with Chapter 2 (cross-chapter contradiction). In the Chapter 1 number list, the diagnostic summarizes it as "dug the drainage ditch half a foot deep, and a very thin layer of white floated up on the wet mud, like mildew frost." At the start of Chapter 2, the same phenomenon is written as if seen for the first time:

   Ch. 1: he clears the drainage ditch, "digging half a foot deep," and sees the white frost layer; Ch. 2, opening: again, "a very thin layer of white turned up on the mud bank where the water had receded... it came off with a fingernail scrape."

   The diagnostic's suggested fix (paragraph numbers follow the diagnostic report): "Either keep Chapter 1 as foreshadowing and cut that half-sentence from Chapter 2, or the other way around. This also ties into a count of three white layers (P69 of Chapter 2 has a separate remark, 'the soil color turned from gray to brown'), so I recommend keeping only the Chapter 1 instance."

The diagnostic also listed 3 open questions for the author to rule on. First, based on the five days of eel removal, water diversion, pond splitting, and seedling raising across Chapters 2 through 4, plus keeping to the old ways, about 7 to 8 days have actually passed, yet Chapter 5 says "the boundary stone has been in for three days." Second, in Chapter 3 Limping Chen says "I'll come help you carry the beam tomorrow," and neither Chapter 4 nor 5 delivers on it or explains the no-show. Third, the diagnostic pointed out that the protagonist already discovers at the end of Chapter 1 that his qi only circulates when he's working, but the start of Chapter 3 reads as if he's discovering it all over again.

At a Glance

Keeping Progression Fantasy Numbers Straight: What You Can Try

1. Write the stat rules into your story bible or character profile before you start. If you plan to write a LitRPG with a status screen, settle the level-up and point-allocation formula, starting attributes, skill ranks, and currency conversion before generating Chapter 1. In this test, DeepSeek set a no-system premise in Chapter 1. While continuing Chapters 2–5, it didn't add status screens so as not to overturn the earlier setup, and after delivering, it asked the author whether a major rewrite was wanted.
2. Try leaving a short closing inventory at the end of each chapter. You don't need to paste the full status screen into every chapter, but you can try noting the end-of-chapter level, remaining money, key item counts, and current day of the week in your chapter outline, giving later chapters a baseline to check against.
3. Try writing fixed rules into your writing guidelines, and bring the previous chapter's numbers along when you continue. In this test, DeepSeek put prices and setting rules into a file called Writing Guidelines and attached it every time it handed a chapter to the writer (though 2 hard contradictions still showed up over five chapters, so a follow-up check is still needed).
4. Watch for fraction words and time words. In this test, the models didn't get plain arithmetic like 5 + 20 = 25 or 1.5 × 12 = 18 wrong. Instead, contradictions or timing questions turned up in narrative phrases like "half a skillet," "Twelve days ago," and how many days the boundary stone had been in. For container capacities and multi-day timelines, you can try jotting down a schedule as you go.
5. Try running Chapter Diagnostics every few chapters, then go through the findings one by one before deciding what to fix. Let the diagnostic scan read-only and list a chapter-by-chapter number inventory and the contradiction spots. Then you check them against the manuscript and decide which to change and how.
6. The page language isn't the prose language, so try spelling out "Write in English" in your request. In this test, DeepSeek V4.1 Flash still produced Chinese prose after getting English instructions on the English page. If your goal is an English novel, you can try adding "Write in English" to your first request to specify the prose language (we didn't separately test this instruction).

FAQ

Q1: Can AI write a LitRPG?

Yes. You can have it write job levels, cultivation tiers, four attributes, skill level-ups, and silver and supplies into the prose. In this test, Gemini 4 Carbon wrote 5 chapters totaling 19,518 words of English, producing status screens and inventory lists around a kiln-cultivation setting. But Chapter Diagnostics still found 4 contradictions in the prose involving cultivation tier, attribute timing, quantity, and dates, so a manual review is still needed.

Q2: Will an AI-written stat sheet contradict itself?

It will. In Gemini 4 Carbon's 5-chapter test, the diagnostic didn't flag any conflict in the status screens themselves, but it found 4 places where the prose clashed with earlier setup. For example, the story established that Body reached 9 only late Thursday night, yet the midnight fight said he moved faster than yesterday's Body 9. And Chapter 1 gave a third-tier cultivation, while in Chapter 5 a supporting character recalls him as only second-tier two weeks earlier. When you check, go through the prose and dialogue too.

Q3: How do I get AI to remember levels and numbers?

You can try writing the starting rules into your story bible or character profile, noting the current level, attributes, skills, money, and timeline in your chapter outline at the end of each chapter, and bringing the previous chapter's numbers along every time you continue. Even so, contradictions still appeared in this test with chapter outlines or Writing Guidelines in place, so we suggest running Chapter Diagnostics every few chapters, going through the findings one by one, and deciding by hand how to fix them.

Q4: Why did the English page produce Chinese prose?

Interface language and English instructions don't always lock the output language. DeepSeek V4.1 Flash wrote in Chinese when imitating Beware Of Chicken on the English page, and also when imitating Pride and Prejudice and Sherlock Holmes. If you want English prose on the English page, you can try explicitly adding "Write in English" to your request (we didn't separately test this instruction).

Q5: How many credits does it take to write 5 chapters?

This test had two steps, writing the prose and running the diagnostic. Gemini 4 Carbon spent 996.60 credits writing five English chapters, or 1,193.71 credits with the diagnostic. DeepSeek V4.1 Flash spent 140.68 credits writing five Chinese chapters, or 186.24 credits with the diagnostic. The former's total was about 6.4 times the latter's.

Q6: Can the diagnostic fix things automatically?

In this test, we told it "Don't change the text," so Chapter Diagnostics only listed the chapter-by-chapter number inventory, the contradiction spots, open questions, and suggested fixes, and left all five chapters untouched. After the diagnostic, both models said they would wait for the author to confirm which spots to change and which option to use before making targeted edits or continuing with Chapter 6.

Wrapping Up

From this 5-chapter test, models can set up a numbers framework and keep the big picture of income and spending straight, but cross-chapter recall, calendar count-backs, and everyday fraction words still slip. Set up your story bible ledger before you start writing, run a read-only Chapter Diagnostics check every few chapters, and let the author decide what to change.

Related Reading:

• AI Novel Continuation Outline Test
• Worldbuilding Story Bible Card: A Real Example Breakdown
• Testing AI Style Imitation of Public-Domain Authors
• Same Prompt, Six Models: A Web Novel Test