Factual Grounding Score

Measures the precision with which a generated claim maps to its source. High fidelity requires a direct, unambiguous link between a specific passage in the retrieved context and the output text.

• Direct Quotation Match: Verbatim or near-verbatim alignment with source text.
• Entailment Verification: The source text logically implies the generated statement.
• Granular Citation: The ability to point to a specific sentence or paragraph, not just a whole document.
• Low Fidelity Example: A summary that introduces a causal relationship not present in the source.

Quantifies the degree of agreement across multiple independent, authoritative sources. A claim supported by a single source has a lower score than one verified by a consensus signal from a diverse set of documents.

• Source Diversity Index: Penalizes over-reliance on a single origin.
• Consensus Threshold: The minimum number of agreeing sources required for a high score.
• Contradiction Detection: A negative signal that actively lowers the score when sources disagree.
• Example: A medical claim confirmed by both the CDC and the WHO receives a higher corroboration metric than one found only on a single blog.

A pre-computed score reflecting the trustworthiness of the evidence source itself, independent of the claim. This is derived from a citation graph analysis, similar to PageRank, and is a critical multiplier in the grounding calculation.

• Graph-Based Reputation: Authority is determined by the quantity and quality of inbound citations from other trusted sources.
• Domain Expertise: A peer-reviewed journal has a higher intrinsic rank than a social media post for scientific claims.
• Trust Discounting: A source's rank is decayed if it has a history of propagating retracted or contradicted information.
• Dynamic Score: Authority is not static; it evolves with the source's continued publication record.

Incorporates a data freshness stamp to evaluate if the evidence is still current. A perfectly grounded statement from a 10-year-old source may be penalized if the knowledge domain has a high rate of change.

• Confidence Decay Function: A mathematical formula that systematically reduces the score as the source ages.
• Staleness Threshold: A hard cutoff where data is considered too old to be reliable and is excluded entirely.
• Domain-Specific Windows: Financial data may have a validity window of minutes, while historical facts may have a window of decades.
• Freshness-Aware Ranking: The grounding score is recalculated based on the most recent version of a source.

Distinguishes between uncertainty from missing knowledge (epistemic) and inherent noise (aleatoric). A well-calibrated grounding score reflects low epistemic uncertainty, meaning the model has sufficient evidence to be confident.

• Expected Calibration Error (ECE): A primary metric for measuring the gap between the model's confidence and its actual accuracy.
• Temperature Scaling: A post-hoc logit calibration technique to align confidence scores with true likelihoods.
• Conformal Prediction: An alternative framework that provides a mathematically guaranteed set of plausible answers instead of a single score.
• High Epistemic Uncertainty: The model's grounding score should be low if the knowledge base lacks information on the topic.

The process of assigning different levels of importance to various pieces of evidence before computing the final score. A direct, empirical measurement in a source is weighted more heavily than an anecdotal mention.

• Evidence Hierarchy: Primary sources (e.g., clinical trial results) are weighted above secondary sources (e.g., a literature review).
• Subjective Logic: A framework that models belief, disbelief, and uncertainty as separate components, allowing for nuanced weighting.
• Negative Weighting: Contradictory evidence is not just ignored; it actively subtracts from the score.
• Contextual Relevance: A passage that directly answers the query is weighted more than a tangentially related one.

The cryptographic foundation for any grounding score. Source attestation embeds a verifiable claim of origin, authorship, and integrity directly into content. Without it, an AI model cannot establish a provenance chain to differentiate a primary source from a re-blogged rumor. Techniques include digital signatures and content hashing that allow a retrieval system to cryptographically confirm that a document has not been tampered with since its creation by a known, trusted entity.

The definitive metric for auditing the quality of a confidence score. ECE partitions a model's predictions into bins based on confidence (e.g., 0-10%, 10-20%) and computes the weighted absolute difference between the average confidence and the actual accuracy within each bin. A perfectly calibrated model has an ECE of zero. Temperature scaling is the most common post-hoc method to minimize ECE on a validation set.

A factual grounding score must decompose uncertainty into its two fundamental types. Epistemic uncertainty is the model's ignorance due to a lack of knowledge or training data—it is reducible by retrieving more evidence. Aleatoric uncertainty is the irreducible noise inherent in the data itself, such as conflicting eyewitness reports. A robust grounding score will be low if epistemic uncertainty is high, signaling a need for more retrieval, but may remain high despite noise if aleatoric uncertainty is the dominant factor.

A critical negative signal for factual grounding. This NLP task identifies when a generated statement is logically inconsistent with a retrieved source or when two authoritative sources conflict. A high contradiction signal against a trusted knowledge base immediately drives the grounding score toward zero. This is a core component of evidence weighting, where a single contradiction from a high-authority source can override multiple low-quality corroborations.

A machine-readable temporal marker that governs the temporal validity window of a fact. A factual grounding score is not static; it decays over time based on a confidence decay function. A statement grounded in a source with a freshness stamp from the last hour will have a higher score than one grounded in a five-year-old document, especially for time-sensitive topics. The staleness threshold defines the point at which a source is excluded entirely from the retrieval set.

A quantitative measure of consensus that directly boosts a factual grounding score. When multiple independent, authoritative sources retrieved from a knowledge base support the same claim, the consensus signal increases the system's trust. The source diversity index refines this by penalizing corroboration from sources that share the same parent origin, preventing a single press release from masquerading as widespread independent verification.