Fiction
Provenance
Every chapter here carries a label saying who wrote it and who went over it afterwards. A tier system nobody has defined in public is just a mood, so these are the definitions, and they are the same ones the site’s own validator enforces — a chapter that does not fit them does not build.
Authorship — exactly one
-
human
Human-written
A person wrote the first draft of every sentence.
-
ai
AI-written
A model wrote the first draft. What a person did to it afterwards is the editing pass below, and nothing else is implied.
Editing passes — stacked on top
An editing pass says who revised the draft and how far it moved. A chapter can carry one AI pass, one human pass, both, or neither. The tiers measure how much of the text changed, not how long it took.
-
light
Light AI edit · Light human edit
Roughly one sentence in ten touched. Word choice, rhythm, a cut line — nothing that changes what happens.
-
medium
Medium AI edit · Medium human edit
Up to about a third of the sentences. Paragraphs restructured, moved or dropped; the shape of a scene may have changed.
-
heavy
Heavy AI edit · Heavy human edit
More than a third. Wholesale rewriting, and the version you are reading may be structured nothing like the draft.
-
de minimis
AI grammar pass
Spelling and grammar tooling only, with no suggestion about phrasing accepted. It never appears beside an AI editing pass, because an editing pass already covers everything a grammar pass would catch.
How a label reads
Labels are listed in the order the work happened: written first, edited after. AI-written · Heavy human edit means a model produced the draft and a person reworked most of it. Human-written · Light AI edit means the opposite, and the AI’s contribution was line-level.
Where both an AI and a person edited at the same tier, both labels are shown and neither is presented as the dominant pass. That is a real state and it is stated plainly rather than rounded to whichever sounds better.
When a model had a hand in a chapter, it is named. Several of the models that appear here are also scored in the benchmark, which is the same judgement applied to a controlled prompt instead of a chapter.
What this is not
These labels are declared, not detected. No reliable detector for AI-written prose exists, and nothing here pretends otherwise: the labels are worth exactly what the person setting them is worth. They are published because a claim on the record can be wrong in a way that silence cannot.