Why one paragraph carries the whole score
The number describes a document; the evidence usually sits in a handful of sentences.
When a whole document is scored, the figure you see is an aggregate. That is convenient and misleading in equal measure: a tenth of the text can decide the result, and the rest of the document never appears in the number at all. Reading segment by segment is slower, and it is the only way to know what was actually measured.
What this page is for: reading a detection score in context. It is not offered as a way past a checker, and no figure or example here should be taken as a promise about what any detector will report; scores move, and services disagree.
Where a document score comes from
- A document score is a summary. Whatever combination rule produced it, most of the text contributes very little to the result.
- Short segments are noisy. A thirty-word paragraph gives a model almost nothing to work with, so its reading swings with phrasing that carries no information about the writer.
- Quoted and formulaic material crowds the scale. Method sections, block quotes and standard descriptions of routine work read uniformly, and uniformity is what these tools look for.
- A few flat patches can move the total. In a long document the aggregate can be driven by five or ten percent of the words.
- Without segments you cannot tell steady prose from a small number of flat passages, and those two cases call for different responses.
- Segment output is not a verdict either. It narrows the question from whether a document is machine-written to what inside it produced the reading.
How to work through a document
- Start with the aggregate and treat it as a flag, not a finding.
- Move to the segments. Find the spans carrying the highest reading; everything else is context.
- Read each flagged span as a reader before reading it as data. Ask what the sentence is doing, and whether its uniformity belongs to the genre or to the writer.
- Separate quotation from composition. Spans that reproduce someone else's wording are not evidence about the person submitting the work.
- Compare unflagged spans by the same author. The contrast, not the absolute figure, is the useful part of the reading.
- Record which spans moved the score, and why you accepted them or set them aside.
Questions people ask
Why does one paragraph raise the score for a long document?
Because the aggregate is dominated by the spans with the clearest signal. A long document made of varied prose can still produce a high total if a small part of it reads flat, and the summarised figure will not tell you which part.
Are paragraph-level readings more accurate?
They are more specific, which is not the same thing. A segment score localises the reading so you can see what produced it, but the signal in a short segment is weak, so the reading for any single paragraph is less stable than the reading for a whole document.
Should quoted material count?
Not as evidence about authorship. Quoted wording comes from somewhere else. If a quoted block sits inside a flagged span, the span is evidence about the quotation, not about the person who included it.
What if the flagged spans are all method or results sections?
That pattern is common and worth naming. Routine descriptions of procedure are formulaic by convention, and convention is what a uniformity measure responds to.
Do I need segment output to make a decision?
You need to know where the reading came from. If a tool cannot show that, the score is only a flag, and flags are not decisions.