How I Make Graded Chinese Stories

Every story is developed and closely edited by our editors before publication. Our production system assists them with research, early drafts, language checks, pinyin, and audio preparation; editors make the creative and publishing decisions.

What the Editors Own

A graded story starts with an editorial premise, not a model response. Editors decide what the story is about, why its events happen, and what the reader should understand or feel by the end.

That ownership continues through the whole story. An editor checks that causes lead to believable consequences, characters remember what has happened, names and objects stay consistent between chapters, and the language register suits both the scene and the intended HSK reader. Revision is not a final polish applied to an otherwise finished draft; it is where the story is shaped.

Automated checks can flag a repeated passage or an above-level word. They cannot decide whether a character's choice feels earned, whether a reveal is satisfying, or whether a difficult word is worth teaching in context. The editor makes those judgments and makes the final publish decision.

Where the Production System Helps

The production system reduces repetitive work around that editorial process. It can help gather research, surface outline alternatives, and produce early drafts that an editor can compare, combine, or reject. Once a direction is chosen, the same system can run deterministic chapter, language, translation, pinyin, and media checks.

It also turns reviewed source material into repeatable exports. Story structure, translations, pinyin data, and media inputs need to stay aligned as a story changes. Keeping those steps reproducible gives editors more time for the parts that require reading and judgment.

This division of labor matters: the tools make iteration faster and catch specific classes of error, while editors remain responsible for the work a reader actually encounters.

What Later Experiments Changed

Our experiments are snapshots, not permanent laws about writing. Two July 2026 screens changed how we interpret some of our older benchmark results.

In one controlled HSK 7 architecture screen, giving the draft system a complete causal solution outline produced a more coherent story than giving it only a brief. Adding a separate per-chapter state card improved some objective diagnostics but made the result more expository and scored worse on craft. That is evidence that useful structure can help and extra structure can hurt; it is not evidence that one outline recipe wins for every genre or level.

In a separate HSK 7 writer screen, GPT-5.6 won the blind craft scoring and was strongest on plot reliability. A closer editorial read found that Fable 5 had the stronger prose and advanced register, but weaker plot coherence. Most importantly, no raw writer output in that screen was ready to publish as-is. Each still needed editorial work on issues such as continuity, register, or how the solution was revealed.

Those results are useful inputs to editing. They do not replace editing or establish a universal ranking of models and prompting methods.

Grading the Language

We use two vocabulary views because they answer different questions.

Running-token unknown rate asks how often a reader encounters an unknown word while moving through the text. It is unknown running-word occurrences / all running-word occurrences; its companion metric, token coverage, reports the known share of that same denominator. If the same unknown word appears five times, all five encounters affect this measure.

Unique-type error rate asks how much of the story's distinct vocabulary pool violates the target level. It is distinct violating words / all distinct words. A word used five times counts once here.

These percentages are not interchangeable. A story can repeat a small set of unknown words and therefore look different by token and type measures. We use both as diagnostics, run deterministic checks against the target vocabulary, and then read the actual passages. Editorial judgment determines whether the wording is clear, natural, and appropriate for the learner; the story still has to pass its release gates.

Pinyin and Audio

Pinyin is generated through a multi-stage language pipeline rather than a character-by-character lookup. Chinese text first has to be segmented into words; phrase pronunciations, polyphonic characters, tone changes, and character-to-syllable alignment then need their own checks. The full process is explained in How to Add Accurate Pinyin to Chinese Text.

Audio has a separate production and review workflow. Every published story includes Mandarin audio checked for pronunciation, pacing, and synchronization with the text. Generation alone is not the acceptance test: the workflow also checks complete chapter coverage, listens to representative passages, and validates the timing data used for highlighting. The technical system is covered in Running Chinese Text-to-Speech on a GPU.

Current Production Snapshot

Checked July 29, 2026. The current generation routes in storybuilder.toml are:

These routes are production settings, not endorsements for every Chinese writing task. We change them when new evidence about language control, coherence, reliability, cost, or editorial quality warrants a change. The dated snapshot above is therefore more useful than treating any model name as a permanent part of the method.

What Makes a Story Ready

A story is ready only when two kinds of evidence agree. An editor has approved the premise, causal logic, character continuity, register, revisions, and final reading experience. Deterministic gates have also passed for structure, language, translations, pinyin, audio, timestamps, and release metadata.

Neither side is sufficient alone. A clean validation report cannot make a weak story compelling, and a strong story cannot ship with broken pinyin, missing audio, or unresolved production errors.

For the deeper evidence and technical systems, see:

• The Chinese LLM benchmark
• The pinyin pipeline
• The Mandarin text-to-speech pipeline