> ## Documentation Index
> Fetch the complete documentation index at: https://docs.scite.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Verification tools

> Check claims and reference lists against the literature, record citation decisions, and format bibliographies through MCP.

The verification tools check scientific writing before anyone relies on it. They work on text from any source: a model's own draft, a user's manuscript, or citations recalled from memory or found by web search. You don't need to have found the sources through Scite.

Requires a Scite premium subscription.

## The verification loop

The server's instructions tell connected models to run this loop on any scientific answer before returning it:

**1. Check the claims**

Call `check_claim` on each substantive claim to pull the verbatim statements for and against it. Drop or flag any claim nothing supports.

**2. Check the citations**

Call `verify_papers` on every DOI or reference about to be cited, wherever it came from.

**3. Record the decisions**

Call `report_citations` once with every source cited and every source considered and dropped, each with a reason and its provenance. Call `citation_report` to show the include/exclude record.

`bibliography` then formats the final reference list from stored metadata rather than from memory.

## check\_claim

Pass up to 20 claims as plain statements. For each one, `check_claim` returns the sentences in the literature closest to it, labeled by source, with the retraction and dispute status of each paper.

```json theme={null}
{
  "name": "check_claim",
  "arguments": {
    "claims": [
      "high-dose diclofenac raises major vascular events by about 40%",
      "oral diclofenac became prescription-only in the UK in January 2015"
    ],
    "alt_phrasings": {
      "high-dose diclofenac raises major vascular events by about 40%": [
        "diclofenac increases cardiovascular risk"
      ]
    }
  }
}
```

### Parameters

| Parameter | Type | Description |
| - | - | - |
| `claims` | string\[] | Required. Up to 20 single statements |
| `dois` | string\[] | Check only against these papers, such as a set you already retrieved. Omit to search the whole corpus |
| `alt_phrasings` | object | Up to 2 paraphrases per claim, keyed by the exact claim string. The search is lexical, so paraphrases catch papers that word the finding differently |
| `limit` | integer | Papers searched per claim. Default `20`, maximum `30` |
| `screen` | boolean | Return only each claim's `summary`, without quotes or papers. Use it to grade many candidate claims cheaply |
| `trend` | boolean | Add `tallyByYear`: citation stance counts bucketed by citing year |
| `filters.type` | string | Keep one Smart Citation stance: `criticism` (contradicting), `supporting`, or `mentioning`. The papers' own passages are always kept |
| `filters.as_of` | integer | Drop papers, citing statements, and editorial notices from after this year |

### What it returns

Each claim gets a result with these fields:

| Field | Contents |
| - | - |
| `status` | `found`, `not_found`, or `error`. `not_found` means no matching statements, not that the claim is false |
| `summary` | Counts to branch on: `n_stating_papers`, `n_supporting`, `n_contrasting`, `has_retraction`, `n_distinct_sources` |
| `matchedStatements` | The matching sentences. `stance: "states"` is the paper's own text (`source` is `fulltext` or `abstract`). Smart Citation snippets carry `stance` of supporting, contrasting, or mentioning and `source: "smart_citation"` |
| `papers` | Every quoted paper, with its `tally` and `editorialNotices` |
| `readNext` | Up to 5 DOIs of papers that state the claim, to pass to `read_fulltext` or `citation_graph` |
| `readNextCiting` | Up to 5 DOIs of citing papers holding the contrasting (first) or supporting passages |
| `tallyByYear` | With `trend: true` only. A rising contrasting share is a signal to read the contrasting passages, not a verdict on the claim |

The `summary` counts are not a verdict. A keyword match isn't entailment, so confirm each quote actually states the claim before relying on it.

> [!TIP]
>
> To get the claim's citation graph in one call, pass the claim to [`citation_graph`](https://docs.scite.ai/mcp/tools/citation-analysis#citation_graph) as `claim` instead of calling both tools.

## verify\_papers

Pass up to 20 references per call, as bare DOIs, references as written, or a mix. Each one comes back `verified` with the matched record, or `could_not_verify` with a `reason`.

```json theme={null}
{
  "name": "verify_papers",
  "arguments": {
    "dois": ["10.1056/nejmoa1200303"],
    "references": [
      "Rieder CL, Maiato H. 2004. Stuck in division or passing through. Developmental Cell 7:637-651."
    ]
  }
}
```

| Parameter | Type | Description |
| - | - | - |
| `dois` | string\[] | DOIs to verify. `https://doi.org/` URLs are accepted |
| `references` | string\[] | References as written, one per string. A printed DOI is checked first; otherwise the text is matched against Scite's records |

A verified reference can also carry:

| Field | Meaning |
| - | - |
| `printedDoi`, `note` | The printed DOI didn't match but the reference text did. The result gives the correct DOI |
| `editorialNotices` | Retraction, withdrawal, or expression of concern, each with `status`, `noticeDoi`, and `date` |
| `predatoryJournal` | `true` when the journal is on Scite's predatory-journal blocklist |
| `contrastingEvidence` | One classifier-flagged contradicting citation from later work: `citingDoi`, `section`, `excerpt`. `excerpt` is `null` when the license withholds the text |

`could_not_verify` means Scite couldn't confirm the paper exists, either in its records or at doi.org. It doesn't mean the paper is fake. Check it before citing.

## report\_citations

`report_citations` records an answer's full source decision set: every source it cited and every source it considered and excluded. Call it once per answer.

```json theme={null}
{
  "name": "report_citations",
  "arguments": {
    "context": "Does high-dose diclofenac raise cardiovascular risk?",
    "citations": [
      {
        "source_ref": "10.1016/s0140-6736(13)60900-9",
        "decision": "cited",
        "source": "scite_mcp",
        "reason_code": "supports"
      },
      {
        "source_ref": "10.1007/s12671-012-0144-z",
        "decision": "excluded",
        "source": "web_search",
        "reason_code": "off_topic",
        "stage": "title_abstract"
      }
    ]
  }
}
```

Each item in `citations` takes:

| Field | Description |
| - | - |
| `source_ref` | Required. A DOI (preferred), URL, or reference string. A bare `[N]` reference number is rejected |
| `decision` | Required. `cited` or `excluded` |
| `source` | Where the source came from: `scite_mcp`, `web_search`, `user_supplied`, or `other` |
| `source_detail` | Names the source when `source` is `other`, e.g. `arxiv` or `model_memory` for a citation recalled rather than retrieved |
| `reason_code` | For excluded: `off_topic`, `retracted`, `contradicted`, `duplicate`, `low_quality`, `superseded`, `out_of_scope`. For cited: e.g. `supports`, `relevant` |
| `reason` | Optional free-text note |
| `stage` | Optional PRISMA screening stage: `title_abstract` or `full_text` |

Pass an optional `answer_id` to group decisions, and `context` for the question or inclusion criteria.

`source` records provenance honestly. Confirming a citation with `verify_papers` doesn't make its source `scite_mcp`.

The response returns `recorded_cited`, `recorded_excluded`, `skipped` (malformed items), and the accepted decisions as `cited[]` and `excluded[]`. When a cited DOI that didn't come from Scite is retracted, withdrawn, or under an expression of concern, the response adds `editorialAlerts` and a `message`.

> [!NOTE]
>
> For organizations with no-retention handling of query text, `reason`, `reason_code`, `source_detail`, `context`, and any non-DOI `source_ref` are not stored. In the Scite Assistant, decisions appear in the answer's fact-check panel and are not stored, so `citation_report` has nothing to summarize there.

## citation\_report

`citation_report` summarizes the decisions recorded with `report_citations` for the current session, or for one `answer_id`. It's read-only.

It returns the identified, screened, included, and excluded counts (`retrieved` is `null` for an answer-scoped report), `excluded_by_reason`, `by_stage`, `by_source`, `checks` (such as decisions missing a reason), `truncated`, and the full `included[]` and `excluded[]` lists. That's the screening funnel a PRISMA flow diagram or a regulatory evidence submission needs.

## bibliography

`bibliography` formats papers you've already found as a reference list built from Scite's stored metadata, so keys and fields are exact.

```json theme={null}
{
  "name": "bibliography",
  "arguments": {
    "dois": ["10.1038/s41586-020-2649-2"],
    "format": "ris"
  }
}
```

| Parameter | Type | Description |
| - | - | - |
| `dois` | string\[] | DOIs to format. Order and de-duplication are preserved |
| `pmids` | string\[] | Up to 50 PubMed IDs, formatted after the DOIs |
| `format` | string | `bibtex` (default), `ris`, `enw` (EndNote), `csv`, or `apa` |

For Zotero, Mendeley, or ReadCube Papers use `ris` or `bibtex`; for EndNote use `enw`. `apa` returns an APA 7 list ready to paste into an answer, covering journal articles only.

The response carries `format`, `found`, `notFound`, and `content`. A DOI that matched only after dropping trailing punctuation is listed in `corrected`.

## Related

* [Citation analysis tools](https://docs.scite.ai/mcp/tools/citation-analysis.md): `citation_graph` and the claim graph
* [Reference check guide](https://docs.scite.ai/guides/reference-check.md): reference checking over HTTP
* [Prompts](https://docs.scite.ai/mcp/prompts.md): the `fact-check-claim` and `verify-bibliography` workflows


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.