Rules and marks
The rulebook holds 92 rules today. Each one names a habit, such as stock vocabulary like “delve” and “tapestry”, or a paragraph that calls the same company “the firm”, “the vendor” and “the provider” in three sentences running.
Every rule carries a pair of examples, one with the tell and one without. When the rulebook is built, the engine runs over each example and records the exact characters it caught. Those are the yellow marks you see, so a mark is always the engine's own output and never a hand-drawn highlight.
Some rules measure a whole paragraph, such as how evenly formal every sentence is. Those have nothing to point at, so they fire without a mark and say “Measured on every paragraph” on the rules page.
3 voice rules never feed the verdict. They exist to stop the rewrite engine from flattening a writer's own voice.
The paragraph verdict
The dial shows a machine share from 0 to 100 for the paragraph at the reading line. Right now that number comes from a simple placeholder: the scores of the rules that fired are added up and mapped onto the scale. It is not calibrated, and the checker says so beside the dial.
The calibrated house model replaces the placeholder in one place when it ships. Its accuracy is being measured and will be published here with the method.
The house model
The model is trained on Compare the Cloud's own archive from before 2022, the year ChatGPT launched, set against AI rewrites of the same article titles. Training on pairs with the same subject means it learns style rather than topic.
It is tuned for UK English business and technology prose. Spellings like “organisation” and dates like 22 September 2026 count as normal, not as tells. On fiction, poetry or text translated from another language it is likely to be wrong, and its error rate there is being measured.
Can't tell
A paragraph under 25 words, or one where no verdict rule fired, reads “can't tell”. The checker would rather say that than guess. Expect it often on short or plain paragraphs, and read it as the honest answer rather than a failure.
A machine-ish mark is a guess about style. People write in stock phrases too, and an edited model draft can read human. Never use a mark on its own to accuse anyone of anything.
Where Jev fits
Jev is a model that judges prose quality. Slop Check asks it plain questions about a paragraph, and only for the few rules where the engine's own measurement lands in a grey zone, such as whether a paragraph keeps swapping in new names for the same company.
Jev is never the human or machine classifier. In a study it flagged about a fifth of paragraphs written by people as machine-made, which is far too many for a verdict about authorship.
Your text
On the checker page, scoring runs in your browser. The text you paste is not uploaded, and a draft is kept only in this browser's local storage so it survives a reload.
Fix it is the one exception. When you press it, the paragraph at the reading line is sent to CTC's server, rewritten by an open model on Cloudflare, checked by Jev to confirm every fact survived, and sent back. Neither the paragraph nor the rewrite is stored. If you accept or reject a rewrite, the rule ids and your choice are kept, never the text.
The API scores text on CTC's server and does not store or log it. The only things kept are counts of which rules fired, with no text attached.