Sixteen AI models each wrote the same short story twice from their own outline: once cold, once with a written craft method. Six judges scored every draft without knowing the model or the condition. New models are added as they ship, each addition a numbered version.
Ranked by score with the method. V3, scored 20 September 2026. Overall score out of 10, the average of thirty blind scores. Silver is the cold draft; gold is the same model and outline with the method. Select a model for every category and every judge.
V4: the launch-day models
Three models shipped on 22 September 2026. Each was run the same day, same protocol, same method version, and scored blind by the same six judges on 23 September. One scored higher with the method. Two scored higher cold.
Ranked by score with the method. Opus 5.5 needed a second outline: the provider's content filter cut its first cold draft five times out of five, so both of its drafts were rerun from a fresh outline to keep the pair honest. The first outline and every cut draft are kept on file.
What the run found
By category
The V3 average across thirteen models, by what the judges were scoring.
How it was run
A result is only worth what its method is worth. Here is ours, including what it cannot tell you.
The protocol
- Each model wrote its own outline. Same brief for all thirteen, so no model was handed a plot that suited another.
- One outline, two drafts. The cold draft got the outline and nothing else. The method draft got the same outline plus the Sound Human method. The method is the only thing that changed.
- No memory, no preferences. Every run went through an API in a fresh context. No account history, no saved instructions, no custom skills.
- Six blind judges. Each story carried only a letter. Judges did not know the model or whether the method was used.
- Five categories, 1 to 10: prose, character, structure, voice and pull. A story's overall score is the average of all thirty scores it received.
- Checked twice. Totals were recomputed independently from the raw scores before publication.
- Versioned. V3 is the thirteen-model field of 20 September 2026. V4 added the three models that shipped on 22 September, run and judged the same way. Each new addition gets the next number; the field is never rerun to flatter the method.
What this does not show
- One story per model per condition. A second story could land differently, and small gaps between neighbouring models are not rankings to bet on.
- Short fiction only. It says nothing about a novel's structure over 300 pages.
- These scores measure craft as judges read it. They are not an AI detector and say nothing about who wrote a text.
- Models change. Every number here is dated. New models are added when they ship, under a new version number.
- The method was tuned on the V3 field. In V4 the two GPT-6 models scored lower with it than without. That is on the page because it happened.
The method is public.
The Sound Human method is a written set of craft instructions given to the model before it drafts. It is taught in full in the free Stop Slop classroom, so anyone can repeat this run.