Corroboration Metric

A computed score reflecting the perceived trustworthiness and expertise of a content source. This is often derived from a Citation Graph analysis, where algorithms like PageRank evaluate the network of citations to determine influence. A claim supported by a high-authority source (e.g., a peer-reviewed journal) receives a greater weight than one from an unverified blog. This component directly addresses the question: 'Is the source known to be reliable?'

A confidence-boosting indicator derived from multiple independent, authoritative sources corroborating the same factual claim. The metric quantifies the Source Diversity Index, which measures the breadth of unique, independent sources. A high consensus signal penalizes over-reliance on a single origin and rewards broad agreement. For example, a scientific fact reported identically by five unaffiliated research labs generates a much stronger signal than the same fact repeated across five syndicated news outlets.

The process of assigning different levels of importance to various corroborating or contradicting sources when calculating a final confidence score. This is not a simple average. A source with a high Source Authority Rank is given more weight. Furthermore, a source that directly refutes a claim triggers a Contradiction Detection penalty, which can drastically lower the overall metric. This component ensures that a single highly credible source can outweigh several low-quality ones.

A defined period during which a piece of information is considered accurate and relevant. The corroboration metric incorporates a Confidence Decay Function that systematically reduces the trust score as data ages. A Data Freshness Stamp is checked against a Staleness Threshold; if the corroborating evidence is too old, it is excluded from the calculation entirely. This ensures that a well-corroborated but obsolete fact (e.g., the number of planets in 2005) does not outrank a newer, correct one.

A metric evaluating how well an AI-generated statement is supported by verifiable evidence from a specific, retrieved knowledge source. It goes beyond source authority to assess the direct link between a claim and its citation. High Attribution Fidelity—the accuracy with which a model cites the correct source passage—is a prerequisite. This component answers the question: 'Does the cited source actually say what the claim asserts?'

A mathematical framework for reasoning under uncertainty that explicitly models belief, disbelief, and uncertainty as separate components of an opinion. Unlike a simple probability, this allows the metric to distinguish between 'confidently false' and 'unknown.' It also formalizes Trust Discounting, a function that reduces the weight of a source's opinion based on a computed distrust factor, preventing unreliable sources from skewing the final corroboration score.

The algorithmic process of assigning different levels of importance to various sources when calculating a final corroboration score. Not all evidence is equal; a peer-reviewed journal carries more weight than a social media post.

• Authority-based weighting: Prioritizes sources with high Source Authority Rank.
• Recency-based weighting: Applies a Confidence Decay Function to older data.
• Statistical weighting: Favors sources with low Aleatoric Uncertainty in their measurements.

A confidence-boosting indicator derived from multiple independent, authoritative sources corroborating the same factual claim. It is the positive output of a high Corroboration Metric.

• Threshold-based: A claim is 'verified' only when the number of agreeing sources exceeds a predefined Staleness Threshold.
• Graph analysis: Uses a Citation Graph to ensure sources are truly independent and not simply citing a single, flawed origin.

An NLP task that identifies logically inconsistent statements across sources, serving as a direct negative input to the Corroboration Metric. It prevents the formation of a false consensus.

• Entailment analysis: Determines if one statement logically follows from another.
• Neutral/Contradiction classification: Standard NLI (Natural Language Inference) labels used to map semantic relationships.
• Conflict resolution: Triggers a higher Epistemic Uncertainty score when contradictions are detected.

A metric that measures the breadth of unique, independent origins supporting a claim. It penalizes over-reliance on a single domain or author, ensuring the Corroboration Metric reflects a genuine consensus rather than a single loud voice.

• Domain uniqueness: Counts distinct root domains in the supporting evidence set.
• Author independence: Analyzes co-authorship networks to ensure sources are not all from the same research group.
• Bias mitigation: A high index directly reduces the risk of systemic bias in the final trust score.

The accuracy with which a generative AI model correctly cites the specific source that supports a claim. A high Corroboration Metric is useless if the model points to the wrong document.

• Precision vs. Recall: Measures if the cited source actually contains the claim (precision) and if all supporting sources are cited (recall).
• Grounding verification: Cross-references the generated text with the Source Attestation of the cited document.
• Provenance Chain integrity: Ensures the citation links to the original source, not a derivative copy.

The primary metric for evaluating if a model's confidence scores are trustworthy. It measures the gap between a model's perceived accuracy (confidence) and its actual accuracy. A perfect Corroboration Metric should drive ECE toward zero.

• Binning: Predictions are grouped by confidence level (e.g., 0-10%, 10-20%).
• Calculation: ECE = weighted average of |accuracy( bin ) - confidence( bin )| across all bins.
• Reliability diagrams: Visual plots used to diagnose calibration errors.