Factual Consistency Check

The system breaks down generated text into individual, verifiable atomic claims—single factual assertions that can be independently checked. Each claim is isolated and mapped to source spans. This granular approach prevents complex sentences with mixed accuracy from passing through undetected.

• Example: 'The CEO, who was appointed in 2019, announced record profits' becomes two claims: [Appointment: 2019] and [Record profits announced]
• Enables precise contradiction detection at the sentence fragment level
• Powers downstream metrics like claim recall and hallucination rate

Each atomic claim is paired with its source evidence and run through a Natural Language Inference (NLI) model. The model classifies the relationship as entailment (supported), contradiction (refuted), or neutral (insufficient evidence).

• Neutral classifications flag unsupported inferences—claims that sound plausible but lack source backing
• Modern implementations use fine-tuned models like DeBERTa-v3 trained on datasets such as ANLI and VitaminC
• Scores aggregate into an overall faithfulness metric for the response

Rather than trusting a single retrieved document, the check requires multiple independent sources to corroborate a claim before it passes. This reduces vulnerability to source-level errors and adversarial content.

• Threshold-based: A claim must be entailed by at least N distinct documents
• Weighted voting: Sources with higher reliability scores carry more weight
• Critical for high-stakes domains like medical and legal applications where single-source errors are unacceptable

The check verifies that claims align with the temporal context of both the query and the source material. A statement that was true in 2020 may be false today—temporal grounding catches these staleness errors.

• Extracts date anchors from claims and compares against source publication dates
• Flags anachronisms: 'The current president' referencing a source from a prior administration
• Integrates with document freshness scoring in the retrieval pipeline

When multiple sources conflict, the consistency check doesn't simply fail—it applies resolution logic to determine the most reliable answer. Strategies include majority voting, source authority weighting, and recency prioritization.

• Conflict types: Direct contradiction, outdated information, domain-specific disagreement
• Unresolvable conflicts trigger uncertainty signaling rather than silent failure
• Feeds into confidence calibration to align output confidence with empirical accuracy

Rather than checking after generation, grounded decoding constrains token probabilities during inference to favor words and phrases explicitly supported by the evidence context. This shifts consistency checking from post-hoc to real-time.

• Manipulates logit distributions to penalize unsupported tokens
• Uses attention alignment to verify generated tokens attend to relevant source spans
• Reduces latency by eliminating a separate verification pass while maintaining factual rigor