Source Credibility Score

A positive signal applied to citations from established, high-trust top-level domains and institutional repositories. This component leverages the inherent credibility of long-standing, verified organizations.

• Top-Tier Domains: .gov, .edu, and .mil receive the highest baseline boost.
• Institutional Repositories: Recognized databases like arXiv, PubMed, and IEEE Xplore are weighted favorably.
• Mechanism: Acts as a prior probability in a Bayesian credibility model, adjusting the score before content is even analyzed.

The application of an author-level metric that measures both the productivity and citation impact of a researcher's publications. This component weights the credibility of a cited work based on the author's proven track record.

• Definition: An author with an h-index of 20 has published 20 papers, each cited at least 20 times.
• Disciplinary Normalization: Scores are normalized against the median for the specific field (e.g., biology vs. computer science).
• Limitation: This is a signal of academic impact, not a guarantee of factual accuracy in a single paper.

A temporal decay function applied to a citation's authority score, prioritizing recently published or updated sources to ensure information freshness. This component prevents outdated, potentially superseded research from dominating results.

• Exponential Decay: The score decreases exponentially based on the time elapsed since the publication date.
• Field-Specific Half-Life: The rate of decay is tuned per domain. A paper on deep learning decays faster than one on Newtonian physics.
• Version Awareness: A recent update or revision to a document resets the decay curve, treating it as a fresh source.

A binary or categorical indicator confirming whether a cited source has undergone formal peer review. This serves as a strong heuristic for academic rigor and methodological soundness.

• Binary Flag: true for peer-reviewed journals and conferences, false for pre-prints and self-published work.
• Categorical Granularity: Advanced models distinguish between single-blind, double-blind, and open peer review.
• Predatory Journal Filter: A classifier is run in parallel to down-weight sources from publications characterized by fraudulent editorial practices, preventing them from receiving this flag.

A verification technique that checks for agreement among multiple independent, high-quality sources to confirm a claim. This component increases confidence through corroboration, penalizing isolated or outlier assertions.

• Corroboration Logic: A claim's credibility increases proportionally to the number of independent, high-scoring sources that support it.
• Independence Check: The system verifies that corroborating sources are not simply citing each other, using Bibliographic Coupling Strength analysis.
• Contrarian Penalty: A claim contradicted by a high-consensus body of evidence receives a severe negative weight.

A dynamically updated registry of academic papers, articles, or datasets that have been officially withdrawn. This component acts as a hard gate, automatically invalidating any AI citation referencing a blacklisted source.

• Dynamic Updates: The blacklist is synchronized in near real-time with sources like the Retraction Watch Database and Crossmark.
• Cascade Invalidation: Citations that heavily rely on a retracted paper are also flagged for review, implementing a Citation Chaining Protocol.
• Reason Tagging: Retractions are tagged with a reason (e.g., data fabrication, ethical violation, honest error) to inform downstream risk assessment.

An algorithmic assessment of a source's importance within a network of citations, analogous to PageRank. Authority is derived from the quantity and quality of inbound links from other credible sources.

• Mechanism: Recursive weight propagation through a directed graph
• Key Differentiator: Evaluates structural prestige, not just content quality
• Example: A paper cited by 50 high-Impact-Factor journals outranks one cited by 500 predatory journals

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).

• Mechanism: A classifier assigns labels: Entailment, Neutral, or Contradiction
• Key Metric: A ratio below 0.5 signals potential hallucination or citation misattribution
• Example: A claim stating 'X causes Y' with a source that only shows correlation would yield a low entailment score

A temporal decay function applied to a citation's authority score, prioritizing recently published or updated sources to ensure information freshness.

• Mechanism: Applies an exponential or linear decay based on the publication date
• Domain Sensitivity: Highly weighted in medicine and technology; lower in historical studies
• Example: A 2024 clinical trial carries significantly more weight than a 2010 study on the same drug

A dynamically updated registry of academic papers, articles, or datasets that have been officially withdrawn. It is used to automatically invalidate any AI citation referencing them.

• Integration: Cross-references databases like Retraction Watch and Crossmark
• Critical Function: Prevents the propagation of fraudulent or erroneous research
• Example: The Wakefield MMR vaccine study is a canonical entry, automatically nullifying any citation that relies on its data

A positive signal applied to citations from established, high-trust domains to reflect their inherent institutional credibility.

• Tier 1 Domains: .gov, .edu, .mil, and recognized institutional repositories (e.g., WHO, CERN)
• Mechanism: A multiplier applied to the base credibility score before other factors are calculated
• Example: A citation from nih.gov receives an automatic baseline boost over an identical claim on a personal blog

A verification technique that checks for agreement among multiple independent, high-quality sources to confirm a claim, increasing confidence through corroboration.

• Mechanism: Clusters semantically similar claims and checks for factual alignment across distinct sources
• Key Principle: A claim verified by three unrelated Tier 1 sources achieves a near-certain confidence score
• Example: A scientific fact reported identically by Nature, Science, and a major university lab is flagged as consensus-verified