Patterns slopscore does not score¶
slopscore matches phrases, sentence shapes, and Markdown block shapes. The patterns on this page sit at the level of the whole document. They show up often in drafts that went through several AI-assisted passes, and also in drafts nobody ran through a model. None of them can be checked with a regex and a word list without flagging good writing too, so slopscore leaves them to a human reader. This page is a checklist for that read, not an authorship test: finding these patterns in a document says nothing about who wrote it.
Each entry gives the pattern, a short invented example, why slopscore does not score it, and how to check for it by hand.
Thesis restated in every section¶
Example. The summary says the program runs in four stages: detect, triage, fix, prevent. Section 2 opens by naming the four stages again, section 3 closes on them, and the FAQ repeats them.
Why it is not scored. Repeating a system's name or a key term in every section is correct
term consistency. The defect is restating the claim, which takes meaning to detect. slopscore
catches only the verbatim form: REDUNDANT_REPEATED_SENTENCE flags a sentence of 15 or more
words that recurs later with at most a word or two changed.
How to check. For each paragraph, ask what fact it adds that no other paragraph states. If the answer is none, delete it. State the thesis once, in the summary.
A wrap-up paragraph closing every section¶
Example. Each section ends on a paragraph that opens "This shows that..." or "Together, these steps..." and carries no number, name, or date.
Why it is not scored. Some closers are worth keeping, and position alone does not separate
them. slopscore flags the most formulaic openers (FORMULAIC_LESSON_FRAME: "the larger lesson
is", "this case illustrates how", "taken together,"; FORMULAIC_IN_CONCLUSION) wherever they
appear, but it does not judge a section's last paragraph as such.
How to check. Read the last paragraph of each section on its own. If it restates the section without a fact, end the section on the paragraph before it.
Framing before evidence¶
Example. A results section opens with two paragraphs on why the approach matters before the first measurement appears in paragraph three.
Why it is not scored. Some sections need context first. slopscore reports evidence density
(genericity, and runs of factless sentences in metadiscourse), but not where in a section the
first fact falls.
How to check. Find the first number, case, or named example in each section. If it is not in the first two sentences, move it up.
Claims larger than their support¶
Example. A heading promises a large dollar figure that the body never derives. A verb such as "demonstrates" rests on a pilot of a few dozen units, most of which were not yet measurable.
Why it is not scored. Checking a derivation means following arithmetic and sources. A rule such as "a figure in a heading must appear again in the body" would flag every document that states a figure once.
How to check. For each headline number, point to the line that computes it and the source of each input. Match the verb to the sample: "found" or "measured" for a pilot, "showed" for a result that held, "demonstrated" only when the sample would survive review.
Lists of abstract nouns¶
Example. "This reduces idle capacity, repair delays, and unnecessary cost." Five-item lists of this kind in consecutive paragraphs.
Why it is not scored. Lists of three are ordinary English. slopscore's parallelism
dimension already treats rule-of-three as a weak tell that needs corroboration; scoring every
noun list would flag most technical writing.
How to check. Keep the item that matters most and give it a number. Cut the others or move them to a table.
One system, several names¶
Example. The same service is called by its project name, an acronym, and a descriptive phrase in different sections.
Why it is not scored. Telling a synonym from a second system needs knowledge of the domain.
How to check. Search for each name. Pick one, define it at first use, and replace the rest.
An ask buried in a modifier¶
Example. "This is an important area for continued investment" where the writer needs two people for two quarters.
Why it is not scored. The request is implied. Nothing in the sentence marks it as an ask.
How to check. Have someone who did not write the document read only the summary and say what is being asked, of whom, and by when. If they cannot, write the ask as a sentence with an owner and a date.
Present tense for something that does not exist yet¶
Example. "The service collects the logs, ranks the causes, and records the fix" in a document whose plan section says the first version has not shipped.
Why it is not scored. Present tense is normal for describing a design. Only the rest of the document shows whether the design has shipped.
How to check. For each capability described in the present tense, check whether it runs today. If it does not, write "will" or "is designed to", and say which parts are manual today.
Rebuttals of positions nobody took¶
Example. "Rather than building a separate service for each team, we extend the existing one" when no reader proposed separate services.
Why it is not scored. The contrast is fine when someone did propose the alternative, and
only the reader knows whether they did. slopscore scores the most formulaic shapes of this move
("not merely X but Y", "it's not X, it's Y") in parallelism.
How to check. For each "rather than" or "instead of", name who proposed the alternative. If nobody did, state what the design does and drop the contrast.
Draft scaffolding¶
Example. TBD in a table column, a TODO in brackets, italic prompts left under headings, a notes section marked for deletion.
Why it is not scored. People leave these in their own drafts, and they mark an unfinished
document, not a generated one. slopscore flags only unfilled template slots ([X date], [$X],
[Company Name]) under RESIDUE_TEMPLATE_SLOT, and assistant placeholders such as [insert
date] under RESIDUE_PLACEHOLDER.
How to check. Search for TBD, TODO, [, and italic lines before anyone else reads the
document. Track the unfinished parts in a separate list.
Corrective disclaimers about sources¶
Example. "Claims about the stone's suspension should be treated as religious tradition rather than as an archaeological explanation." "The guidance does not establish that the park has a contamination problem."
Why it is not scored. The October 2026 revision of WP:AISIGNS calls these the newest hedging pattern, nearly absent before mid-2026. They matched none of 90 Wikipedia articles tagged as suspected AI text from July 2026 on, and "rather than" alone is more common in human text than AI text in every older corpus slopscore has. A rule would be unmeasured.
How to check. For each "should be treated as" or "does not establish", ask who suggested the stronger reading. If nobody did, cut the sentence or state what the source does say.
Vague association¶
Example. "He is associated with the city orchestra. He later became associated with music education in the region." Four or five such links in one short biography.
Why it is not scored. WP:AICONNECT lists it among the most common 2026 signs, but one "is associated with" is ordinary English: it appeared in 14 human and 7 AI documents across the corpora. The tell is density, and no corpus has enough 2026 text to set a threshold.
How to check. Replace each "associated with" with the relation it hides: taught at, founded, played in, was CEO of. If you cannot name the relation, the source probably does not support it.
Synonym cycling¶
Example. The same person is "the protagonist", then "the key player", then "the eponymous character" in successive sentences.
Why it is not scored. Detecting it needs to know that three phrases name one referent. Italian and other school traditions also teach writers to avoid repeating a word, so it is common in non-native English.
How to check. Pick one name for each person or thing and use it every time.
Premise stacking¶
Example. The point is stated, restated in other words, and stated a third time before the paragraph reaches its claim.
Why it is not scored. Repetition of an idea in different words needs a model of meaning.
slopscore catches only the verbatim form (REDUNDANT_REPEATED_SENTENCE) and near-identical
adjacent sentences (REDUNDANT_ADJACENT_PAIR).
How to check. Read each paragraph's first two sentences. If the second only restates the first, delete one.