Source-Output Divergence Metric

The foundational mechanism that computes the cosine similarity or Euclidean distance between the high-dimensional embedding of the source text and the embedding of the AI's generated claim. A low similarity score indicates a high degree of divergence.

• Uses transformer-based models to generate contextual embeddings.
• Compares sentence-level and document-level vectors.
• Establishes a threshold beyond which a claim is flagged for review.

A critical sub-component that uses Natural Language Inference (NLI) to determine if the source text logically entails the generated claim. This goes beyond semantic similarity to check for directional logical support.

• Classifies relationships as entailment, contradiction, or neutral.
• Detects when an AI over-extrapolates a finding, e.g., a study on mice being cited for a claim about humans.
• Provides a probabilistic Factual Entailment Ratio.

A specific detection module designed to identify when a generated claim broadens the scope of a source's conclusion without evidence. This is a primary failure mode in AI-generated summaries.

• Example: A source states 'Vitamin C may reduce cold duration,' but the AI output claims 'Vitamin C prevents colds.'
• Uses contrastive analysis between the source's conclusion section and the generated claim.
• Flags speculative language in the output that is absent in the source.

Measures the loss of critical qualifiers like 'may,' 'suggests,' or 'in a specific population' when moving from source to output. A high divergence score often correlates with the stripping of academic hedging.

• Tracks the presence of epistemic markers in the source.
• Penalizes the AI output for transforming a tentative finding into a definitive statement.
• Ensures the generated text maintains the source's original level of certainty.

Extends the metric beyond text to evaluate divergence when an AI interprets a chart, table, or image from a source. This checks for misreading of visual data.

• Compares the AI's textual summary of a chart against the chart's underlying data table.
• Detects axis mislabeling or trend misrepresentation in generated descriptions.
• Flags instances where a specific data point is cited but its value is hallucinated.

A composite index that categorizes the detected divergence by its potential impact, ranging from minor stylistic paraphrasing to critical factual inversion.

• Level 1 (Negligible): Synonym substitution with no semantic shift.
• Level 3 (Major): Omission of a key limiting condition.
• Level 5 (Critical): A generated claim that directly contradicts the source's primary finding.

The calculated probability that a cited source document logically supports or entails a specific claim made in AI-generated text. This is determined through Natural Language Inference (NLI) models, which classify the relationship between a premise (the source text) and a hypothesis (the AI's claim) as entailment, contradiction, or neutral. A low entailment ratio is a primary trigger for a high divergence metric.

A composite metric that quantifies the degree of semantic and factual correspondence between a specific AI-generated statement and the content of its cited source. Unlike simple keyword overlap, this score uses dense vector embeddings to measure conceptual similarity and cross-referencing against a knowledge graph to verify factual consistency. It serves as the direct inverse of the Source-Output Divergence Metric.

The process of identifying when a cited source's content has been updated or altered post-citation, potentially invalidating the original evidence. This is critical because a divergence metric calculated at time t may become obsolete if the source webpage changes. Drift detection systems continuously monitor cited URLs and re-evaluate the divergence score, flagging instances where the AI's claim is no longer supported by the live source.

A classification of how precisely a citation points to its evidence, ranging from a full document to a specific passage, sentence, or data point. A divergence metric is far more meaningful when paired with granular attribution. A claim citing an entire 50-page report has a higher risk of unsupported extrapolation than one citing a specific paragraph. High granularity enables precise divergence measurement at the sentence level.

A predictive score estimating the likelihood that a generated statement is a hallucination. It is calculated by analyzing multiple signals, including the absence of supporting citations, high internal model uncertainty (logit entropy), and a high Source-Output Divergence Metric. The divergence metric is a critical input feature for this index, transforming a semantic gap into a quantifiable risk signal for downstream filtering.

A verification technique that checks for agreement among multiple independent, high-quality sources to confirm a claim. If a single source is cited but its content diverges from the AI's output, a cross-reference check against other authoritative sources can determine whether the AI's extrapolation was valid or a hallucination. Consensus strengthens confidence even when a single source's alignment is imperfect.