Qwen3.6 35B wrote long-form prose locally on Ollama for Windows. Fourteen deterministic checks covered headings, chronology, object state, event order, exact text, dialogue count, and length. Every attempt and each check is shown as measured.
| attempt | total | load | rate | checks | what happened |
|---|---|---|---|---|---|
| Cold initial | 51.83 s | 24.41 s | 71.75 tok/s | 10/14 | 1,503 words; four contract failures. |
| Warm identical | 27.66 s | 0.29 s | 71.22 tok/s | 10/14 | Byte-for-byte identical to the cold response. |
| Warm repair 1 | 28.71 s | 0.29 s | 73.22 tok/s | 11/14 | Best run; missed by two words, one time, and one verb. |
| Clean repair 2 | 21.31 s | 0.32 s | 73.06 tok/s | 10/14 | Narrower feedback regressed to 1,146 words. |
The first draft created a flooding lunar railway station, carried a sealed envelope through rising water, and gave two characters recognizable voices. It read like a real story.
These fourteen checks measure explicit instruction control, not literary merit: headings, chronology, object state, event order, exact text, dialogue count, and length. Voice, imagery, causality, subtext, rhythm, and the ending sit on a separate blinded editorial worksheet.
The first targeted repair reached 11 of 14. It ran 1,602 words — two over the ceiling — omitted one required timestamp, and changed an exact required verb.
A narrower second repair returned to 10 of 14 at 1,146 words. One additional run is kept but excluded from the speed comparison because a supervised ComfyUI process restarted during the protocol and occupied the GPU.
The prose was usable as a draft. Long-form fiction, game narrative, screenplays, and branded content all carry state that polished language can quietly contradict, so the explicit contracts are checked directly.