Evidence Weighting

Assigns a pre-computed trust score to each source based on its position within a citation graph. Algorithms like PageRank evaluate the network of citations between documents, boosting the weight of sources frequently cited by other authoritative entities.

• Mechanism: Graph centrality analysis on the provenance chain
• Effect: A peer-reviewed journal carries more weight than an anonymous forum post
• Key Metric: Authority score propagated through the citation graph

Boosts the weight of a claim when multiple independent, authoritative sources converge on the same factual statement. The corroboration metric quantifies the degree of agreement across disparate origins.

• Positive Signal: Three unaffiliated research labs reporting identical findings
• Negative Signal: Contradiction detection flags logically inconsistent statements
• Key Metric: Source diversity index penalizes over-reliance on a single origin

Applies a confidence decay function to systematically reduce evidential weight as data ages. A data freshness stamp provides the timestamp, while a staleness threshold defines when weight reaches zero.

• Example: Financial market data decays in minutes; medical guidelines decay in years
• Mechanism: Exponential or linear decay applied to the temporal validity window
• Integration: Freshness-aware ranking incorporates decay directly into retrieval scoring

Reduces the weight of a source's evidence proportionally to a computed distrust factor. This prevents unreliable or adversarial sources from skewing the final confidence calculation.

• Framework: Subjective logic models belief, disbelief, and uncertainty as separate opinion components
• Application: A source with a history of retractions receives a permanent weight penalty
• Guardrail: Prevents Sybil attacks where many low-quality sources attempt to fabricate consensus

Modulates source weight based on the historical calibration of its claims. If a source's past confidence scores align with actual correctness (measured by Expected Calibration Error), its future evidence receives higher weight.

• Technique: Temperature scaling adjusts raw logits to produce well-calibrated probabilities
• Metric: Low ECE indicates a source's confidence scores reliably predict correctness
• Outcome: Well-calibrated sources are amplified; overconfident sources are penalized

Weights evidence based on the verifiable integrity of its lineage. A content integrity chain uses cryptographic hashes to link sequential document versions, ensuring no tampering occurred.

• Signal: An unbroken provenance chain with source attestation increases evidential weight
• Risk: A broken or opaque data lineage triggers automatic weight reduction
• Implementation: Cryptographic verification of every transformation from origin to current state

A quantitative metric, typically expressed as a probability or percentage, assigned by an AI model to indicate the likelihood that its generated output is factually correct. Evidence weighting directly influences this score by modulating the contribution of each source.

• Raw model logits are transformed into calibrated probabilities
• Scores below a staleness threshold may trigger retrieval of fresher sources
• High-quality evidence weighting produces scores that align with Expected Calibration Error (ECE) benchmarks

A computed score reflecting the perceived trustworthiness and expertise of a content source, often derived from graph analysis of citations and reputation. Evidence weighting multiplies factual signals by this authority coefficient.

• Built on citation graph algorithms similar to PageRank
• High-authority sources receive greater weight in consensus calculations
• Domain expertise is assessed per-topic, not globally
• Authority can decay if a source's calibration drift exceeds acceptable bounds

A quantitative measure of the degree to which evidence from disparate, independent sources supports a single statement. This metric serves as a multiplier in evidence weighting formulas, boosting confidence when multiple authoritative sources converge.

• Requires a high source diversity index to avoid echo-chamber effects
• Contradiction detection acts as a negative signal that reduces the metric
• Works in tandem with consensus signal computation
• Penalizes over-reliance on a single origin, even if highly authoritative

A mathematical formula that systematically reduces the confidence score of a piece of information as it ages. Evidence weighting integrates this temporal dimension to ensure stale data doesn't outweigh fresh, verified claims.

• Common implementations include exponential decay and linear ramp-down
• Each source may have its own temporal validity window
• A data freshness stamp is required as input to the function
• Critical for domains with rapidly evolving knowledge, such as medicine and law

The accuracy with which a generative AI model correctly cites the specific source document or passage that supports a claim. Evidence weighting depends on precise attribution to trace which sources contributed to a final confidence score.

• Measured by comparing generated citations against a provenance chain
• Low fidelity triggers a trust discounting penalty on the source
• Essential for auditability in regulated industries
• Works alongside reference density as a heuristic for content rigor

A mathematical framework for reasoning under uncertainty that explicitly models belief, disbelief, and uncertainty as separate components of an opinion. Evidence weighting maps naturally to this framework, where each source contributes a weighted opinion.

• Opinions are represented as triples: (belief, disbelief, uncertainty)
• Trust discounting reduces the weight of unreliable sources
• Consensus operators combine multiple weighted opinions into a single judgment
• Provides formal guarantees that single-score methods cannot offer