We ran 360 rounds of continuation. Models fail long fiction in exactly three ways

We have now run 360 rounds of novel continuation across nine models and two books — an 8.9M-character Chinese fantasy epic and the palace-intrigue classic Empresses in the Palace. Same protocol every time: continue from a fixed anchor, 20 consecutive rounds, each round’s output appended back into a fixed 16k-token window, then double-blind review with two randomized letter mappings. Every long-run failure we observed fits one of three patterns.

Pattern one: the restart loop

The model ignores everything generated so far and restarts from the original text’s ending — every single round. Gemini 3.1 Pro did this for 20 out of 20 rounds in the fantasy run: twenty phrasings of the same opening instant, a story that never takes its second step.

The root cause was ours, not the model’s. Our 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”. Gemini read “reference” as “not canon” and skipped them all. Three ablations pinned it: old wording, 20/20 restarts; label with no explanation at all, still roughly 6 of 8 rounds restarting; explanation rewritten to state the passages are canonical events that must be treated exactly like original prose — zero restarts in 28 rounds across both books. One sentence of prompt wording was the entire difference between disqualification and mid-table.

Pattern two: the mid-run freeze

The model advances normally, then locks onto its own recent output and repeats it verbatim. Grok 4.5 froze in the middle windows of the fantasy run — one stretch repeats the same two paragraphs three times word for word — then partially recovered. On the palace novel it froze harder: both blind reviewers independently wrote “plot rewind” in their notes, and across twenty rounds the story clock never left the afternoon of the inciting incident. Kimi K2.6 showed a milder version, copying whole passages from its own earlier rounds.

This is the classic autoregressive fixed point: as the context fills with the model’s own text, imitating itself becomes easier than advancing. It is also why one-shot writing benchmarks miss it entirely — the loop needs rounds to build up.

Pattern three: the ending rewind

The strangest one, and it hit the best prose mimic in the field. GPT-5.6 Terra’s early and middle windows were, by both reviewers’ judgment, the closest thing to the original author anyone produced — it even reproduced micro-habits like tilde-marked onomatopoeia and the author’s trademark comma rhythm. Then in rounds 18 through 20 it rewound the entire plot back to the anchor point and replayed chapter one, reusing lines from its own round-2 output. Texture stayed perfect; the story folded back on itself.

The lesson we take from it: surface mimicry and long-range plot coherence are separate capabilities. A model can hold an author’s voice flawlessly while losing track of where the story is. On the palace novel, for what it’s worth, Terra never rewound — it took first-tier placement with a zero-incident run.

What survived twenty rounds

DeepSeek V4 Flash won the fantasy run outright — the only system both reviewers called “improving as it goes”, still opening new plot arcs at round 20. Then it dropped to sixth-seventh on the palace novel, where its plain-spoken instincts fight the author’s ornate register, and its sibling V4 Pro — criticized in fantasy for writing “like a finer-penned author” — took first place, because a finer pen is exactly what that book is written with. GLM 5.2 was the only model in the top three of both genres, with one persistent quirk: half-width quotation marks in Chinese prose, in every run, on both books.

No universal ranking survives contact with a second genre. If your app locks you to one model for a whole book, that is a real cost; per-paragraph model switching on continuation is not a luxury feature, it is how you route around each model’s failure pattern before it locks in.

Follow-up, July 28, 2026: the freeze, one model generation later

Kimi K2.6 — the milder mid-run freeze case above — got a successor, and we ran Kimi K3 through the identical protocol: three 20-round chains per book, mixed with the archived nine systems into a twelve-system pool under fresh anonymous mappings, ranked by two reviewers who didn’t know a new model was present. The freeze numbers first. K2.6’s fantasy chain once measured a 97.6% cross-round twelve-gram repetition spike at round nine, under the same prompt wording the new chains ran on. K3 peaked at 0.5% across all 140 supplementary rounds; no round came near our 30% alert line, and neither reviewer flagged a single verbatim loop on any K3 chain. The old K2.6 chains sitting in the same shuffle got re-flagged by both reviewers for the same old freeze.

The re-ranking moved too: K3’s chains took first and second on the fantasy epic, pushing DeepSeek V4 Flash to third, and first on the palace novel — the first model in this series to reach the top two of both genres. Two caveats, as always: this round ran through our agent-based test harness rather than direct API calls (sampling parameters not under our control; the direction matches the July 18 paired run over Moonshot’s official API), and it is still one book and one anchor per genre. What we take from it: the taxonomy outlives the instances. The restart loop still lives in old prompt wording — which, for the record, never tripped K3 in sixty rounds — and K3 grew fingerprints of its own, closing palace-novel scenes with aphorisms the original author never writes and mixing half-width quotes into Chinese dialogue. Failure modes are per-generation observations, not permanent character traits. Re-test on upgrade.