Attribution Granularity Level

The most coarse-grained form of citation, where a claim is attributed to an entire document without specifying a location. This is the default behavior for many large language models.

• Scope: Cites the full paper, article, or webpage
• Verifiability: Low; requires manual scanning of the entire source
• Use case: General background references where precision is not critical
• Limitation: Fails to meet evidentiary standards for specific factual claims

A mid-resolution citation that points to a specific section, paragraph, or block of text within a source document. This is the current standard for retrieval-augmented generation systems.

• Scope: Cites a contiguous block of text (typically 3-10 sentences)
• Verifiability: Moderate; narrows search to a specific passage
• Mechanism: Often achieved through chunk-level retrieval with metadata tracking
• Advantage: Balances precision with computational efficiency in RAG pipelines

A high-resolution citation that maps a specific claim to an exact sentence or two within the source. This requires advanced natural language inference and source-tracking infrastructure.

• Scope: Cites one to two specific sentences
• Verifiability: High; enables rapid human or automated verification
• Technical requirement: Demands sentence-level chunking and entailment scoring
• Example: "The model achieved 94.3% accuracy" → linked to the exact results sentence in the source paper

The finest-grained attribution level, linking a claim directly to a specific cell in a table, a figure, or a structured data field. This is the gold standard for scientific and financial applications.

• Scope: Cites a specific numeric value, statistic, or structured datum
• Verifiability: Maximum; eliminates ambiguity entirely
• Enabling technologies: Knowledge graph grounding, structured data extraction, and cell-level provenance tracking
• Example: "Q3 revenue was $12.4B" → linked to row 17, column C of the source financial table

A composite attribution pattern where a single claim is supported by multiple non-contiguous spans across one or more source documents. This reflects how complex conclusions synthesize evidence.

• Scope: Cites multiple sentences or data points across different locations
• Verifiability: High but complex; requires cross-referencing
• Mechanism: Uses citation chaining and cross-reference consensus algorithms
• Challenge: Must clearly indicate which part of the claim each span supports to avoid conflation

A scoring principle where the confidence in a citation's accuracy decreases as the attribution granularity becomes coarser. This is a core input to Attribution Confidence Intervals.

• Principle: Document-level citations carry higher uncertainty than sentence-level citations
• Application: Used in Claim-Source Alignment Scores to weight evidence strength
• Formula: Confidence ∝ 1 / (scope of cited text)
• Impact: Directly influences Hallucination Risk Index calculations and downstream trust metrics

A quantitative metric evaluating the trustworthiness of a cited source based on multiple signals:

• Author expertise: H-index, institutional affiliation, publication history
• Domain authority: .gov, .edu, and established institutional repositories receive inherent boosts
• Historical accuracy: Track record of citations that withstand subsequent verification
• Peer-review status: Whether the source has undergone formal academic scrutiny

This score serves as a foundational input for higher-level citation integrity calculations.

A composite metric quantifying the degree of semantic and factual correspondence between an AI-generated statement and its cited source. Key components include:

• Semantic Relevancy Vector: High-dimensional embeddings measuring contextual alignment
• Factual Entailment Ratio: Probability that the source logically supports the specific claim
• Source-Output Divergence Metric: Flags potential misinterpretations or unsupported extrapolations

Low alignment scores trigger automatic re-verification or citation removal.

A statistical range expressing the certainty that a specific claim originates from a given source. This accounts for:

• Ambiguities in attribution: When multiple sources contain similar information
• Model uncertainty: Internal confidence signals from the generating model
• Granularity mismatch: When a claim is more specific than the citation's granularity level

Wide confidence intervals indicate the need for more precise attribution or additional corroborating sources.

The process of identifying when a cited source's content has been updated or altered post-citation, potentially invalidating the original evidence. Mechanisms include:

• Reference Provenance Hash: Cryptographic fingerprint of content at time of citation
• Continuous monitoring: Automated re-crawling of cited URLs and document versions
• Change impact analysis: Assessing whether modifications affect the supported claim

Drift detection is critical for maintaining evidence chain integrity over time.

A measure of the completeness and logical validity of the path from an AI's output claim back through its citations to foundational, verifiable data. Key verification steps:

• Citation Chaining Protocol: Recursively tracing citations to original primary sources
• Cross-Reference Consensus: Checking agreement among multiple independent sources
• Primary Source Priority: Weighting direct, first-hand accounts over secondary interpretations

Broken chains indicate missing links or misrepresented evidence.

A hierarchical categorization system ranking sources by editorial rigor and authority:

• Tier 1: Primary research, peer-reviewed journals, official government data
• Tier 2: Established industry publications, institutional reports, verified expert analysis
• Tier 3: News media, corporate blogs, conference proceedings
• Tier 4: Social media, personal websites, unverified user-generated content

Attribution granularity requirements tighten as source tier decreases, demanding more specific citations for lower-tier sources.