Citation Necessity Score

The baseline filter that classifies widely accepted facts as citation-exempt. Claims like 'Water boils at 100°C at sea level' or 'The Earth orbits the Sun' fall below the necessity threshold. The algorithm evaluates cultural universality, domain consensus, and temporal stability—a fact must be uncontested across multiple authoritative knowledge bases and stable over decades to qualify. This prevents citation clutter while ensuring non-obvious claims always require sourcing.

Every factual claim is categorized along a spectrum of empirical testability. Claims about measurable phenomena, statistical data, historical events, or specific research findings receive high necessity scores. The system distinguishes between:

• Directly verifiable: 'Tesla delivered 1.81 million vehicles in 2023'
• Indirectly verifiable: 'Employee satisfaction correlates with productivity'
• Inherently unverifiable: Subjective opinions or speculative forecasts Only the first two categories trigger mandatory citation requirements.

Citation necessity is not uniform across contexts. The algorithm applies field-dependent thresholds that reflect disciplinary norms:

• Scientific/Medical claims: Near-absolute requirement for peer-reviewed sources
• Legal assertions: Demand primary statute or case law citations
• Journalistic reporting: Requires named sources or official records
• Engineering specifications: Must reference technical documentation or standards bodies A claim about quantum computing in a legal brief faces different necessity criteria than the same claim in a physics paper.

Claims that deviate from established consensus trigger elevated necessity scores through statistical surprisal analysis. The system compares assertions against its knowledge base of accepted facts. A statement like 'Smoking causes lung cancer' has low surprisal and moderate necessity; 'Vitamin D deficiency cures autoimmune disorders' has high surprisal and demands rigorous citation. This mechanism catches extraordinary claims that might otherwise slip through generic filters.

Citation requirements intensify as information ages out of the recency window. A claim about current market conditions demands a citation from the last 24-48 hours; a claim about 19th-century history may accept sources decades old. The decay function is domain-calibrated:

• Fast-decay fields: Technology, finance, current events (hours to days)
• Slow-decay fields: Mathematics, classical literature, fundamental physics (years to decades) This ensures time-sensitive assertions never rely on stale evidence.

The system monitors for active scholarly or public disagreement around specific claims. When a topic has documented conflicting evidence or ongoing debate, the necessity score escalates dramatically. Signals include:

• Contradictory findings across reputable sources
• Retraction notices or corrections on related papers
• Wikipedia edit wars or dispute templates on relevant articles
• Formal challenges in academic literature Claims in disputed territory require multiple, balanced citations rather than a single source.

A predictive score estimating the likelihood that a generated statement is a fabrication unsupported by evidence. The index synthesizes multiple signals:

• Citation absence: Claims made without any supporting reference
• Internal model uncertainty: Token-level probability distributions indicating low confidence
• Source-output divergence: Semantic distance between the cited source and the generated text
• Domain complexity: Higher risk in specialized fields like medicine or law When the Hallucination Risk Index exceeds a threshold, the Citation Necessity Score mandates additional verification steps before the output is presented to users.

A verification technique that checks for agreement among multiple independent sources to confirm a claim. Rather than relying on a single citation, this method:

• Identifies corroborating evidence across distinct domains and authors
• Increases confidence through triangulation of independent observations
• Flags claims supported by only one source for additional scrutiny
• Applies a consensus threshold: claims must be verified by a minimum number of high-quality, unaffiliated sources This approach mirrors academic peer review and journalistic fact-checking standards, reducing the risk of propagating errors from a single flawed source.

A classification of how precisely a citation points to its evidence within a source document. Granularity levels range from coarse to fine:

• Document-level: Cites the entire paper, article, or webpage
• Section-level: References a specific chapter or section
• Passage-level: Points to a particular paragraph or set of sentences
• Sentence-level: Directly links to the exact supporting statement
• Data-point-level: Cites a specific figure, table cell, or datum Higher granularity enables more efficient verification and demonstrates rigorous sourcing. The Citation Necessity Score may require finer granularity for extraordinary or controversial claims.

A temporal decay function applied to a citation's authority score to ensure information freshness. The weight adjusts based on:

• Publication date: Newer sources receive higher weight
• Domain half-life: Fast-moving fields like technology have steeper decay curves than slower-moving fields like philosophy
• Update history: Recently revised or maintained sources retain authority longer
• Version tracking: Detection of whether a source has been superseded by newer research This mechanism prevents AI systems from citing outdated or obsolete information, ensuring that the Citation Necessity Score reflects the current state of knowledge.