# How results work

> Scores, summaries, matched sources and matched text - reached deliberately, never dumped.

Results are **raw** - scores, counts, sources, and matched text. Your agent interprets, aggregates, and explains.

## The four levels

| Level | Returns |
|---|---|
| **`scores`** | The `Plagiarism score`, the `AI score`, and word counts. Nothing else |
| **`summary`** *(default)* | The scores, the word counts, and **how many sources matched** - not the sources themselves |
| **`detail`** | **The matched sources**, paged and filterable |
| **`full`** | Every match with its exact text on both sides, paged |

<Note>
  **Nothing is withheld at any level.** The default is narrow so a long document does not arrive all at once - a widely-copied document can match hundreds of sources. *"What did this match?"* needs `detail`; `summary` gives the count, not the list.
</Note>

Scores are percentages, **0-100 to one decimal place** - `98.5`.

## Results, never the document

**What comes back is the flagged parts, never the document itself** - matches, AI content, AI phrases and `Excluded text`, not the text between them. The document stays in Copyleaks and cannot be retrieved through the connector.

Every completed scan carries a link to its report: **`Open this scan in Copyleaks`**. No finding is held back from the results - the report is where the whole marked-up document lives.

## Match types

Strongest first. The same labels your report and scan settings use.

| Type | Means |
|---|---|
| `Exact match` | Identical text |
| `Minor changes` | Small edits to vocabulary, grammar, or verb tense |
| `Paraphrased` | Restructured sentences and rewritten ideas |

## Where matches come from

| Source | Means |
|---|---|
| `Web` | The open internet |
| `Shared data hub` | Documents submitted by the Copyleaks community |
| `Private cloud hub` | Your own private library |
| `This Batch` | Your other files scanned in the same submission |

`This Batch` matches appear only when `Cross-check documents` is on, and count toward the score like any other match. A score carried **entirely** by `This Batch` usually means the same document was sent twice.

Any source may also carry the **`AI Source Match`** tag, meaning that source is itself AI-generated. It is a tag on a source, never a source type.

## AI results - two layers

**AI content** - passages identified as AI-written. **AI phrases** - specific phrases within that content. Phrases are always inside AI content, never a separate category.

Each AI phrase carries a **frequency**: how many times more often that phrase appears in AI writing than in human writing. A phrase at `220x` turns up 220 times more often in AI text.

<Warning>
  **Higher frequency means more distinctive of AI writing - it does not mean the passage is more likely to be AI.**
</Warning>

## `Cheat Detection`

Where a plagiarism scan found them, hidden characters, white ink and text manipulation come back as flagged spans with their exact text and position - in `copyleaks_get_results` and in document order from `copyleaks_get_flagged_segments`.

## `Excluded text` is not `Excluded sources`

- **`Excluded text`** - content your profile's `Exclusions` left out of the score: references, quotes, citations, table of contents, headings, headers and footers, code comments.
- **`Excluded sources`** - sources *you* removed from the score afterwards.