Check-Worthiness Detection

The claim must be objectively resolvable against a corpus of established evidence. This excludes subjective opinions, predictions, or aesthetic judgments.

• Objective Claims: "The unemployment rate fell to 3.5%" is verifiable; "The policy is unfair" is not.
• Evidence Anchoring: The claim must reference entities and events that exist in structured knowledge bases or document corpora.
• Temporal Grounding: Verifiable claims are typically anchored to a specific time point, enabling evidence retrieval from the correct temporal window.

The claim must carry sufficient societal, financial, or safety-related weight to justify computational and human review resources.

• Harm Potential: Claims about medical treatments or election integrity receive highest priority due to real-world consequences.
• Virality Signals: Social engagement metrics and propagation velocity serve as proxies for public impact.
• Entity Prominence: Claims involving high-profile entities (heads of state, major corporations) are prioritized over obscure subjects.

Complex sentences are decomposed into atomic sub-claims that can be independently verified. A single sentence may contain multiple check-worthy assertions.

• Claim Decomposition: "The CEO resigned after profits fell 20%" yields two claims: (1) the CEO resigned, (2) profits fell 20%.
• Predicate Extraction: Each atomic claim contains exactly one predicate-argument structure.
• Granularity Control: Overly granular decomposition can lose context; the system balances atomicity with semantic coherence.

Previously verified claims are deprioritized to avoid redundant computation. The system maintains a cache of resolved claims with their veracity labels.

• Semantic Hashing: Claims are encoded into dense vectors for near-duplicate detection against the verified claim store.
• Temporal Decay: A claim may become check-worthy again if new evidence emerges that could alter the previous verdict.
• Paraphrase Resilience: The detection model must recognize semantically equivalent claims expressed with different syntax or vocabulary.

Claims containing quantitative values, percentages, or comparative statistics are flagged with high priority due to their objective verifiability and potential for manipulation.

• Magnitude Sensitivity: A claim of a "50% increase" is more check-worthy than a "slight uptick" due to the precision of the assertion.
• Comparative Claims: Statements comparing two entities ("X outperformed Y") require dual evidence retrieval.
• Statistical Framing: The model detects when statistics are presented without proper context (base rate neglect, cherry-picked timeframes).

Claims that assert a direct causal link between events are inherently check-worthy because causality is difficult to establish and frequently misrepresented.

• Explicit Causal Language: Phrases like "led to," "caused by," and "resulted in" trigger high-priority classification.
• Correlation vs. Causation: The system flags claims that imply causation from mere correlation, a common vector for misinformation.
• Mechanism Verification: Causal claims require evidence of a plausible mechanism, not just temporal precedence.