Text Coherence & Authorship Analysis

AI analysis of text for internal consistency, claim coherence, and authorship patterns.

The verify-text-coherence pipeline analyses a piece of text — an article, statement, document extract, or any prose content — for internal logical consistency, factual coherence, and authorship patterns. It extracts verifiable claims from the text, checks whether those claims are mutually consistent and plausible given known facts, and assesses whether stylistic and structural features are consistent with the claimed authorship or publication context.

This pipeline is useful for a wide range of verification tasks: checking whether a purported statement contains internally contradictory claims, assessing whether a text has been synthetically generated or substantially modified from an original, and flagging articles that mix accurate and inaccurate information in ways designed to mislead. It is intentionally fast — the AI does not perform live web searches or external source lookups at this tier. For full fact-checking with researcher crowd review, see verify-news-article.

Pipeline Details

Property	Value
Pipeline ID	`verify-text-coherence`
Tier	1 — AI-Only
Estimated Duration	~1–3 min
Credit Cost	700 credits per run
Crowd Stages	No
Files Required	Yes — enriched text (TXT, MD, or HTML) or enriched PDF

Required Inputs

Files

One enriched text or PDF file. The enrichment step extracts the textual content and populates metadata.wordCount, metadata.charCount, and (for PDFs) metadata.author and metadata.creationDate. The file must have enrichmentStatus: "completed" before the run is created. See Files & Enrichment for details.

Context Keys

All context keys are optional. Providing them significantly improves the quality of the consistency analysis — especially claimed_author, which allows the AI to assess stylometric consistency against the named author's known writing patterns.

Key	Required	Description
`claimed_author`	Optional	Claimed author name (e.g. `"Dr. Anna Weber"`). Used for stylometric cross-referencing if the author has a public writing corpus.
`claimed_publication_date`	Optional	Claimed publication date in ISO 8601 format (e.g. `"2024-09-01"`). Cross-referenced against any timestamps embedded in the file and against the plausibility of claimed events.
`topic`	Optional	Subject matter or topic for context-aware analysis (e.g. `"vaccine safety"`, `"EU trade policy"`). Helps the AI assess whether named entities, statistics, and references are plausible within the topic domain.

Stages

Claim Extraction (AI)

Identifies verifiable claims within the text: factual assertions, named entities (people, organisations, locations), dates, statistics, and attributed quotations. The AI produces a structured list of claims, each tagged with its type and the degree to which it is specific and checkable. This list forms the input for the consistency analysis stage.

Consistency Analysis (AI)

Checks the internal consistency of the extracted claims against each other and against plausible world knowledge. The AI evaluates whether the timeline of events is coherent, whether named entities are used consistently throughout the text, and whether numerical claims (statistics, vote counts, measurements) are internally consistent. It also assesses whether the overall framing and rhetorical structure are consistent with the claimed author or publication context, and flags passages that appear stylistically distinct from the rest of the document.

Synthetic Text Screening (AI)

A dedicated AI stage scores the likelihood that the text was generated or substantially rewritten by an LLM. It looks for stylometric tells — sentence-length uniformity, overuse of generic discourse markers, repetitive phrasing, and the absence of idiosyncratic personal voice or specific anecdotes. The stage writes a syntheticLikelihood score (0–100) and a list of syntheticIndicators to its artifacts — see Synthetic Data Detection.

Verdict Synthesis (AI)

Produces the final coherence verdict, weighting the number and severity of consistency failures, and the synthetic-screening result, against the length and complexity of the document. A short document with a single minor inconsistency may still receive a high-confidence authentic verdict; a document riddled with internal contradictions and implausible claims will receive a manipulated or inconclusive verdict, and one with strong LLM-generation evidence will receive a synthetic verdict, each with a detailed explanation of the specific factors found.

Starting a Run

curl -X POST https://api.crowdee.ai/v2/projects/{projectId}/verification-runs \
  -H "X-API-Key: crw_YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "pipeline_id": "verify-text-coherence",
    "file_ids": ["{fileId}"],
    "context": {
      "claimed_author": "Dr. Anna Weber",
      "topic": "vaccine safety",
      "claimed_publication_date": "2024-09-01"
    }
  }'

Example Response

{
  "id": "run_01j9x...",
  "status": "completed",
  "verdict": "inconclusive",
  "confidence": 64,
  "pipeline_id": "verify-text-coherence",
  "scorecard": {
    "claim_extraction": "completed",
    "internal_consistency": "warn",
    "stylometric_consistency": "pass"
  },
  "explanation": "The document contains 14 verifiable claims. Two statistical claims are mutually inconsistent: a stated 78% efficacy rate in paragraph 3 contradicts the 61% figure cited in paragraph 7 with no explanation for the discrepancy. Stylometric analysis is consistent with the claimed author. Verdict is inconclusive pending external source verification."
}

For full fact-checking with live source assessment and crowd researcher review — suitable for newsroom use cases where external claim verification is required — see verify-news-article.