Source Reliability Scoring

Combines multiple authority signals into a single, actionable trust metric. The model ingests diverse data points—from citation integrity and factual consistency to adversarial robustness—and weights them based on temporal decay and contextual relevance.

• Composite Scoring: Merges domain history, author expertise, and content quality.
• Temporal Weighting: Recent accuracy violations impact the score more heavily than historical errors.
• Contextual Adjustment: A source's reliability can vary by topic; a domain might be authoritative on science but unreliable on politics.

Evaluates a source's past claims against a corpus of established evidence to build a track record. This involves automated claim detection and veracity prediction on a publisher's historical output to identify patterns of misinformation.

• Evidence Retrieval: Matches past claims against trusted knowledge bases.
• Precision/Recall Analysis: Measures how often the source's factual assertions were verified as true.
• Drift Detection: Identifies sources that have shifted from reliable to unreliable over time, flagging potential domain takeovers or editorial changes.

Assesses reliability not just by content, but by how information spreads. Sources that consistently originate claims later debunked as disinformation or that amplify known low-credibility domains receive lower scores.

• Cascade Analysis: Tracks how a source's claims propagate through social and citation networks.
• Co-Citation Mapping: Identifies clusters of sources that frequently cite the same unreliable references.
• Bot & Coordination Detection: Flags domains exhibiting inorganic amplification patterns indicative of coordinated inauthentic behavior.

Rewards sources that demonstrate algorithmic explainability in their own processes. Domains that publish clear editorial standards, correction policies, and source attribution protocols are scored higher.

• Correction Velocity: Measures how quickly and visibly a source corrects errors.
• Attribution Density: Analyzes the frequency and quality of primary source citations.
• Schema Compliance: Checks for structured data like ClaimReview markup, signaling a commitment to fact-checking standards.

Proactively probes sources for vulnerability to manipulation. This involves simulating adversarial robustness testing scenarios to see if a domain's content can be easily poisoned or if it republishes synthetic media without detection.

• Synthetic Media Detection: Scans a domain's multimedia content for AI-generated artifacts.
• Plagiarism & Duplication: Identifies domains that rely on content farming and low-value aggregation.
• Ownership Obfuscation: Flags sources that hide their ownership or impersonate legitimate entities, a common disinformation detection signal.

Shifts from domain-level scoring to entity-level precision. Using Named Entity Disambiguation (NED) and Relation Extraction, the system maps the specific authors, organizations, and sources a domain relies on to their own independent trust scores.

• Author-Level Scoring: Evaluates the track record of individual journalists or contributors.
• Source Provenance: Assesses the reliability of the primary sources a domain cites, not just the domain itself.
• Knowledge Graph Grounding: Anchors the scoring model to a deterministic knowledge graph to prevent circular logic in trust calculations.

Dynamic scoring models that continuously update a domain's trustworthiness based on historical accuracy. Unlike static blocklists, these systems apply recursive weighting—a source's past veracity directly influences the credibility assigned to its future claims. Key mechanisms include:

• Bayesian prior updates after each fact-check event
• Temporal decay to prioritize recent accuracy over historical performance
• Propagation of trust scores through citation networks

The machine learning task of classifying a claim as true, false, or mixed based on aggregated evidence and source reliability signals. Modern veracity prediction models operate as ensemble classifiers, combining:

• Stance detection outputs from evidence documents
• Source reliability scores as a prior probability
• Linguistic features indicating hedging or certainty

The output is a calibrated probability rather than a binary label, enabling nuanced trust assessments.

The algorithmic ordering of retrieved documents by their relevance and probative value before final veracity judgment. Effective ranking prevents garbage-in, garbage-out failures in fact-checking pipelines. Core ranking signals include:

• Semantic similarity between claim and evidence passage
• Authority score of the evidence source itself
• Temporal proximity to the claim's timeframe
• Factual consistency with established knowledge bases

Composite models that aggregate multiple authority and quality signals into a single actionable trust metric. These algorithms ingest diverse inputs:

• Domain-level historical accuracy rates
• Author expertise and credential verification
• Citation graph centrality and peer endorsement
• Transparency indicators like ClaimReview markup adoption

The aggregation function often uses weighted geometric means to ensure a single low signal can significantly depress the overall score.

Algorithmic evaluation of the quality, relevance, and trustworthiness of sources cited by an AI or publisher. This is the downstream consumer of source reliability scores. When a generated text cites a domain, the citation integrity scorer:

• Retrieves the cited source's reliability score
• Verifies the citation actually supports the claim (via textual entailment)
• Checks for citation distortion or quote fabrication
• Flags hollow citations that reference authoritative domains but misrepresent their content

Capturing the complete, auditable chain of data transformations from raw source to final AI output. For source reliability scoring, lineage tracking ensures that a domain's score is always traceable to specific fact-check events. Key components:

• Immutable logs of every claim-verification pair
• Cryptographic hashing of evidence snapshots to prevent retroactive editing
• Provenance metadata using W3C PROV standards
• Enables auditors to replay and validate any score computation