Attribution Fidelity

Measures whether a citation points to the specific sentence or paragraph that directly supports a claim, not just the general document. High fidelity requires that a statement like 'Q3 revenue grew 12%' links to the exact table or sentence containing that figure, not the entire earnings report.

• Passage-level precision: Citation must resolve to the minimal evidential unit
• Contrast with topical relevance: A document about revenue isn't sufficient; the exact data point must be present
• Failure mode: Vague citations that force users to re-read entire source documents to verify claims

Uses Natural Language Inference (NLI) to formally verify that a cited passage logically entails the generated claim. The system checks whether a hypothesis (the generated statement) can be inferred from the premise (the cited source text) with a classification of entailment, contradiction, or neutral.

• Entailment: The source text logically implies the generated statement
• Contradiction: The source text directly refutes the generated statement
• Neutral: The source text is topically related but provides no direct evidence

Evaluation datasets constructed by automatically annotating documents with fine-grained citation labels. Tools like Attribute and AttributionBench create ground-truth mappings between statements and source passages, enabling systematic measurement of attribution fidelity across different model architectures.

• Automated annotation: Uses NLI models to label statement-passage pairs at scale
• Cross-document evaluation: Tests ability to attribute claims when evidence spans multiple sources
• Model comparison: Benchmarks reveal that even advanced models often cite irrelevant passages with high confidence

Distinguishes between two dimensions of attribution quality. Extractiveness measures how much of a generated statement is directly copied from the source. Faithfulness measures whether the meaning is preserved, even when paraphrased. High attribution fidelity demands faithfulness, not verbatim extraction.

• Faithful paraphrase: 'The patient presented with tachycardia' from source 'heart rate was 110 bpm'
• Unfaithful extraction: Copying text verbatim but misrepresenting context or omitting qualifiers
• Trade-off: Overly extractive models may plagiarize; overly abstractive models may hallucinate

Quantitative metrics adapted from information retrieval to evaluate attribution fidelity. Citation Recall measures the proportion of generated claims that have at least one supporting citation. Citation Precision measures the proportion of provided citations that genuinely support their associated claim.

• Recall formula: (Number of cited claims with support) / (Total number of verifiable claims)
• Precision formula: (Number of citations with entailment) / (Total number of citations provided)
• F1 Score: Harmonic mean balancing both dimensions for a single quality metric

Evaluates whether multiple citations to the same source entity or event maintain logical coherence across a generated text. A system with high attribution fidelity will not cite one passage claiming 'the CEO was appointed in 2019' and another claiming 'the CEO has served since 2021' without acknowledging the discrepancy.

• Temporal consistency: Dates and sequences must align across citations
• Entity resolution: Same entity must be referenced consistently across multiple source passages
• Contradiction detection: Systems must flag when cited sources conflict with each other

An automated evaluation process that measures the degree to which a generated summary or statement aligns with the facts presented in a source document. It penalizes contradictions and hallucinations by comparing generated claims against reference passages.

• Uses Natural Language Inference (NLI) models to detect entailment vs. contradiction
• Critical for evaluating RAG system outputs at scale
• Complements attribution fidelity by verifying that cited sources actually support the claim

The documented history of the origin, custody, and transformations of a piece of data. It provides a verifiable chain of custody essential for establishing content trustworthiness.

• Tracks data lineage from creation through every modification
• Enables auditors to trace any generated claim back to its original source
• Forms the evidentiary backbone that attribution fidelity metrics depend upon

A minimal, self-contained, and indivisible piece of information expressed in a single sentence. It serves as the fundamental unit for fine-grained factual verification and decomposition.

• Enables precise attribution mapping at the sentence level
• Used by FActScore and similar metrics to evaluate factual precision
• Breaking content into atomic facts allows pinpoint identification of unsupported claims

A fine-grained evaluation metric that decomposes a generated biography into atomic facts and verifies each one independently against a trusted knowledge source like Wikipedia.

• Calculates a factual precision score by dividing verified facts by total facts
• Directly measures the type of attribution fidelity that matters in production
• Developed by researchers at the University of Washington and Meta AI

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.

• Reduces hallucination rates without external retrieval systems
• Each verification question targets a specific factual claim
• Improves attribution fidelity by forcing explicit self-audit of generated statements

A Schema.org structured data markup used by fact-checkers to publish the verdict of a specific claim. It enables search engines to surface fact-check summaries directly in results.

• Provides a standardized format for claim-level attribution
• Binds a specific claim to its review verdict and source
• Used by Google News, Bing, and other platforms to highlight verified information