Source Reliability Score

This signal evaluates a source's track record of providing factually correct information over time. It is computed by comparing past outputs from a domain against verified ground-truth datasets.

• Mechanism: Incremental Bayesian updating where each verified claim increases the prior weight, while corrections or retractions apply a penalty factor.
• Application: A peer-reviewed journal with a 99.7% correction-free rate receives a higher weight than a blog with frequent factual updates.
• Key Metric: Precision-Recall curve of historical claims against a golden dataset.

This static or semi-static signal quantifies the topical expertise of a source based on its recognized standing in a specific field. It prevents a high-authority financial site from being used as evidence for medical queries.

• Mechanism: Composite score derived from topical citation graphs, author credentials, and institutional affiliation.
• Application: A medical board's guidelines are assigned high authority for health queries but are excluded from legal contract analysis.
• Key Metric: Topical PageRank or HITS algorithm applied to a domain-specific citation graph.

This temporal signal applies a decay function to a source's reliability based on the age of the content and the volatility of the subject matter. It ensures that time-sensitive queries prioritize recent publications.

• Mechanism: Exponential decay function where the half-life is dynamically set by the topic's rate of change (e.g., 6 months for technology, 10 years for historical facts).
• Application: A 2019 article on COVID-19 treatments is heavily decayed, while a 2019 article on Shakespearean sonnets retains its full weight.
• Key Metric: Time elapsed since publication vs. topic-specific half-life.

This signal measures how well a source's claims align with a consensus derived from other independent, high-reliability sources. It acts as a real-time corroboration mechanism.

• Mechanism: Natural Language Inference (NLI) models check for entailment between a candidate claim and a set of claims from trusted seed sources.
• Application: A single report making an outlier claim is flagged for low consistency, while a claim echoed by multiple authoritative outlets is boosted.
• Key Metric: Mean entailment probability score against a consensus cluster of documents.

This signal evaluates the trustworthiness of the data's journey from origin to ingestion. It penalizes sources with broken or opaque chains of custody, which are more susceptible to tampering or silent corruption.

• Mechanism: Cryptographic verification of a data lineage log, checking for immutable hashes and attested transformations at each stage of the pipeline.
• Application: A financial filing with a verifiable SEC EDGAR lineage scores perfectly, while an anonymously uploaded PDF with no chain of custody is heavily penalized.
• Key Metric: Percentage of the data lineage chain with valid cryptographic attestations.

This signal adjusts the reliability score by detecting and accounting for known political, commercial, or ideological biases that may skew factual presentation. It does not penalize bias itself but calibrates the weight based on the query's need for strict neutrality.

• Mechanism: Classifier models trained on Media Bias/Fact Check (MBFC) and AllSides datasets to assign a bias vector, which is then used to adjust the final score based on the query's objectivity requirement.
• Application: For a query on election results, a source with a known partisan bias is down-weighted in favor of a non-partisan statistical agency.
• Key Metric: Cosine similarity between the source's bias vector and the query's required neutrality vector.