Claim-Source Alignment Score

Measures the cosine similarity between the vector embeddings of the generated claim and the source text. This foundational layer uses models like Sentence-BERT to determine if the two texts are discussing the same topic at a high level.

• High Similarity: The claim and source share a core subject.
• Low Similarity: A red flag indicating a potential topic drift or a completely irrelevant citation.
• Limitation: High semantic similarity does not guarantee factual accuracy; it only confirms topical alignment.

A classification task that determines the logical relationship between a source (premise) and a claim (hypothesis). The model assigns a probability to three classes: Entailment, Contradiction, or Neutral.

• Entailment: The source text logically implies the claim is true.
• Contradiction: The source text directly refutes the claim.
• Neutral: The source provides related information but does not confirm or deny the claim. This is the core engine for detecting factual entailment and is far more precise than semantic similarity alone.

Calculates the Jaccard similarity between the sets of named entities (people, organizations, locations, dates) extracted from both the claim and the source.

• High Overlap: Confirms that both texts are referencing the same specific actors and objects.
• Entity Hallucination: A claim mentioning an entity not present in the source is a critical failure, often indicating a factual hallucination.
• Weighting: Core entities like a study's lead author are weighted more heavily than peripheral location mentions.

A specialized sub-module that extracts and compares quantitative values, percentages, and statistical findings. This process uses SpaCy rule-based matchers and transformer-based question-answering models.

• Exact Match: A direct numerical correspondence is the strongest signal.
• Tolerance Band: Minor rounding differences are accepted.
• Critical Failure: A claim stating 'revenue increased by 20%' when the source states '15%' results in a severe score penalty, as numerical inaccuracy is a high-confidence indicator of a citation error.

An explicit binary classifier trained to identify when a claim directly opposes the cited source. This goes beyond NLI's 'Neutral' class to actively search for polar opposition.

• Example: The claim states 'The treatment was effective,' but the source concludes 'The treatment showed no statistically significant benefit.'
• Methodology: Uses a fine-tuned RoBERTa model on a specialized dataset of scientific contradictions.
• Impact: A detected contradiction immediately floors the alignment score to near-zero, overriding other positive signals.

Evaluates if the claim overgeneralizes a finding that was specific to a narrow context in the source. This metric penalizes scope expansion.

• Source Scope: 'The algorithm outperforms baselines on ImageNet classification.'
• Overgeneralized Claim: 'The algorithm is the superior solution for all computer vision tasks.'
• Mechanism: Uses a fine-tuned entailment model to detect when the claim's domain of validity is a superset of the source's stated domain, flagging an unsupported extrapolation.

The calculated probability that a cited source document logically supports or entails a specific claim made in AI-generated text, determined through natural language inference (NLI). This metric goes beyond surface-level semantic similarity to assess logical consequence.

• Uses transformer-based NLI models fine-tuned on datasets like MNLI and FEVER
• Outputs a probability score between 0 (contradiction) and 1 (entailment)
• Critical for distinguishing between a source that is merely topically related and one that actually proves the claim

A high-dimensional embedding that mathematically represents the contextual meaning of a source document to calculate its topical alignment with a specific AI-generated claim. Unlike keyword matching, this vector captures latent semantic relationships.

• Generated using sentence-transformer models like all-mpnet-base-v2
• Cosine similarity between the claim vector and source vector forms the basis of alignment scoring
• Enables detection of paraphrased support where exact terminology differs

A measurement of the semantic distance between the content of a cited source and the AI's generated text, flagging potential misinterpretations, unsupported extrapolations, or outright fabrications.

• High divergence with high confidence indicates a likely hallucination
• Low divergence with high entailment confirms strong alignment
• Implemented using cross-encoders for fine-grained textual comparison

A classification of how precisely a citation points to its evidence, ranging from a full document to a specific passage, sentence, or data point within the source. Finer granularity enables more accurate alignment scoring.

• Document-level: Citation points to the entire paper or article
• Passage-level: Citation identifies a specific paragraph or section
• Sentence-level: Direct quote or specific claim reference
• Data-point-level: Citation links to a specific table cell or figure

A verification technique that checks for agreement among multiple independent, high-quality sources to confirm a claim, increasing confidence through corroboration. This method reduces reliance on any single potentially flawed source.

• Requires a minimum of 2-3 independent sources for high-confidence claims
• Weighted by each source's individual Source Credibility Score
• Particularly valuable for controversial or emerging scientific claims

A measure of the completeness and logical validity of the path from an AI's output claim back through its citations to the foundational, verifiable data. A broken chain—where a secondary source misrepresents a primary source—invalidates the alignment.

• Validates each link in the citation chain recursively
• Detects citation drift where sources have been updated post-citation
• Uses Reference Provenance Hashes to immutably verify source content at time of access