ClaimReview

The exact statement being fact-checked. This must be a short, verbatim quote of the claim as it appeared in public discourse, not a paraphrase or summary.

• Type: Text
• Best Practice: Keep it under 75 characters for optimal display in search results.
• Example: "The Earth is flat and surrounded by an ice wall."
• Critical Rule: Never editorialize the claim. Use the original wording to avoid introducing bias.

The verdict assigned to the claim, structured as a Rating type. This is the core output that search engines use to display visual fact-check labels.

• Type: Rating
• Common Values: True, False, Mostly True, Mostly False, Mixture, Unproven
• Example: A reviewRating with ratingValue of "False" and alternateName of "Pants on Fire"
• Note: Use bestRating and worstRating to define your scale if using numeric values.

The publisher of the fact-check, which must be an established organization recognized for journalistic integrity. This is a critical trust signal for search engines.

• Type: Organization
• Required Properties: name, url
• Eligibility: Only verified signatories of the International Fact-Checking Network (IFCN) or equivalent bodies are eligible for rich results.
• Example: {"@type": "Organization", "name": "Snopes", "url": "https://www.snopes.com"}

The entity or creative work that contains the claim. This contextualizes the claim by linking it to its source, such as a news article, social media post, or speech.

• Type: CreativeWork or MediaObject
• Use Cases: Link to a specific news article, a Facebook post, or a video segment.
• Example: A NewsArticle with a datePublished and url pointing to the original source of the false claim.
• Why It Matters: Establishes provenance and allows search engines to cluster related fact-checks.

The canonical URL of the fact-check article itself. This must be the permanent, dedicated page where the full analysis and methodology are published.

• Type: URL
• Requirement: The page must contain the full text of the fact-check, not just a summary or embedded markup.
• Best Practice: Use a static, non-paginated URL to ensure the markup remains stable.
• Example: https://www.factcheck.org/2024/01/debunking-viral-mars-claim/

The publication date of the fact-check in ISO 8601 format. This temporal signal is essential for search engines to determine the freshness and relevance of the verdict.

• Type: Date
• Format: YYYY-MM-DD or full ISO 8601 datetime.
• Critical Rule: This must reflect the date the fact-check was published, not the date the original claim was made.
• Example: "2024-07-15"

An automated evaluation process that measures the degree to which a generated summary or statement aligns with the facts presented in a source document. Factual consistency scoring penalizes contradictions and hallucinations by comparing generated text against a ground-truth reference. Common approaches include:

• NLI-based scoring: Using Natural Language Inference models to detect entailment vs. contradiction
• Atomic fact decomposition: Breaking text into minimal claims for independent verification
• Token-level probability analysis: Identifying low-confidence spans that signal confabulation

The documented history of the origin, custody, and transformations of a piece of data, providing a verifiable chain of custody essential for establishing content trustworthiness. Source provenance answers the question 'Where did this claim come from?' by cryptographically binding metadata about:

• Original publication date and author
• Subsequent edits and modifications
• Chain of custody through content pipelines
• Digital signatures via standards like C2PA (Content Credentials)

A minimal, self-contained, and indivisible piece of information expressed in a single sentence, used as the fundamental unit for fine-grained factual verification. Atomic facts are the building blocks of fact-checking pipelines:

• Each fact makes exactly one verifiable assertion
• Facts are independently checked against a knowledge base
• The FActScore metric uses atomic fact decomposition to calculate factual precision
• ClaimReview verdicts often target individual atomic facts within larger claims

A prompting technique where a language model first drafts a response, then generates a series of independent fact-checking questions to systematically verify and correct its own initial output. CoVe implements a self-critique loop:

• Draft phase: Model generates initial response
• Question generation: Model creates verification questions for each factual claim
• Independent verification: Each question is answered without access to the original draft
• Correction pass: Inconsistencies are resolved in a final, verified output

A metric evaluating how accurately a generated statement's citations point to the specific source passages that directly support it. Attribution fidelity ensures that references are not just relevant but precisely evidential. Key dimensions include:

• Granularity: Does the citation point to the exact sentence or paragraph?
• Entailment: Does the cited text logically support the claim?
• Specificity: Is the source authoritative for this particular claim?
• High attribution fidelity is critical for ClaimReview implementations where verdicts must link to precise evidence

The process of aligning a model's predicted probability of correctness with its actual empirical accuracy. A well-calibrated model means a 90% confidence score genuinely reflects a 90% chance of being right. Confidence calibration techniques include:

• Temperature scaling: Adjusting output probabilities post-training
• Conformal prediction: Producing statistically rigorous prediction sets with guaranteed coverage
• Semantic entropy: Clustering outputs by meaning before measuring uncertainty
• Poor calibration undermines ClaimReview by producing verdicts with unreliable certainty signals