We had six models continue the same novel for 20 rounds each, then ranked them blind

The test book is an 8.9-million-character Chinese fantasy webnovel with a very particular fingerprint: short sentences, fragmented paragraphs (35 characters on average), onomatopoeia on its own line, 16% of the text inside dialogue, and roughly one simile per thousand characters. We picked a battle scene at the 55% mark as the anchor, then had six models continue from there — 20 rounds each, every round's output appended back into a fixed 16k-token window that gradually squeezes the original prose out. That setup deliberately mirrors what happens in a real reader app: the more you continue, the more of the context is the AI's own writing.

The lineup: DeepSeek V4 Flash, DeepSeek V4 Pro, Claude Opus 4.8, Gemini 3.1 Pro, Qwen 3.7 Max, GLM 5.2. Same prompt, same temperature, 120 generations total. Scoring ran on two tracks: computable structure metrics (sentence-length distribution, dialogue ratio, simile density), and a double-blind review — outputs anonymized under two different random letter mappings, ranked by two reviewers unaware of each other.

Final ranking

Across both mappings, the top three and the last place matched exactly. Both reviewers independently marked the #4/#5 call low-confidence — the argument is whether “consistently okay” beats “brilliant middle, corrupted ending”, which is a taste weighting, so we report a tie.

Two times the metrics lied

Qwen 3.7 Max should have won on paper: mean sentence length 32.1 versus the original’s 32.1, dialogue ratio 18.1% versus 16.1%, lowest cliché-simile density in the field. The blind review placed it fourth-fifth. The evidence was specific: every window contained smell descriptions — “the acrid burn stinging the nostrils”, “a cool faint fragrance” — in a novel that goes 8.9 million characters without describing a single smell. Sentence statistics cannot measure “wrote something the original never writes”.

The opposite failure: our dialogue detector scored GLM 5.2 at zero. It actually wrote plenty of dialogue — 108 quote pairs, a real ratio of 13.5% — all in half-width quotation marks that the full-width-only regex ignored. The blind reviewers, without seeing any stats, independently caught the same thing and docked it for surface inauthenticity: a Chinese webnovel written entirely in half-width quotes reads wrong at a glance.

The eerie Gemini bug: one opening, rewritten 20 times

Each individual Gemini round looked fine. Reading all 20 in sequence was unsettling: every round re-describes the same instant — “the two dive into the blood array, ripples spread across the cocoon of light” — twenty phrasings of a story that never takes its second step.

The root cause wasn’t model capability. Our continuation context labels which passages are original prose and which are earlier AI continuations, and the explanation line said AI passages were “for plot continuity reference”. DeepSeek read that the way we meant it. Gemini read it as “these passages aren’t canon” — and restarted from the original text’s ending every single round.

Three ablations pinned it down. Old wording: 20/20 regressions. Deleting the explanation entirely, keeping only the labels: roughly 6 of 8 rounds still regressed — an unexplained label defaults, for Gemini, to “probably not canon”. Rewriting it as two explicit statements — AI passages are canonical events that already happened, continue from the very last paragraph, and only anchor style to the original-prose sections — brought it to 0/20 with normal plot advancement. Wording robustness has to be designed for the model most likely to misread it, and the explanation cannot be omitted: no explanation is only marginally better than a bad one.

Two weaknesses every model shared

First, nobody learned the fragmented paragraph rhythm. The original averages 35 characters per paragraph, often one sentence alone; the six models ranged from 62 to 94 — roughly double. Same story for sentence-length burstiness: the original’s coefficient of variation is 0.84, all six models landed between 0.46 and 0.59. Uniform sentence length remains the deepest machine fingerprint we keep hitting across experiments.

Second, scene anchoring drifts collectively. The anchor scene sits inside a blood-cloud forbidden-ground array; twenty rounds later, all six models had independently wandered into forests, valleys, even “sunlight dappling through leaves”. Style instructions govern diction; they do not govern world state — that belongs to a different mechanism (lore entries, state tracking), not the style prompt.

What this means in practice

If you mostly read Chinese webnovels, the data supports a cheap conclusion: DeepSeek V4 Flash took first place in style fidelity while sitting in one of the lowest price tiers of the six. V4 Pro writes sharper dialogue at the cost of denser narration. Claude and Qwen write well — but “writes well” and “reads like the original author” are different products; they fit scenarios where you want the AI to have its own pen, like transformative fanfic, rather than seamless continuation.

All six models plug into Foreverse with your own key, and you can switch models per paragraph while continuing. The label-wording fix from this experiment already shipped — you don’t need to rediscover that bug.