Temporal Consistency

Every fact must be explicitly bound to a validity interval with a defined start and optional end time. This prevents a statement about a person's job title from 2019 from conflicting with their current role.

• Uses ISO 8601 time representations for machine-readability
• Supports open-ended intervals for ongoing states (e.g., "is CEO since 2020")
• Enables temporal reasoning queries like "Who was CFO during Q3 2022?"
• Critical for knowledge graph integrity when entities have evolving attributes

A system exhibits temporal consistency when no two facts about the same entity and attribute assert conflicting values for overlapping time periods. A person cannot be both "employed" and "retired" simultaneously.

• Requires constraint validation against temporal logic rules
• Detected via automated SHACL shapes with temporal operators
• Violations indicate stale data, extraction errors, or missing interval metadata
• Foundational for factual consistency scoring in RAG pipelines

Facts must be compared at compatible levels of temporal precision. A statement granular to a year ("Revenue was $10M in 2023") cannot directly contradict one granular to a quarter ("Revenue was $3M in Q1 2023") without explicit sub-interval reasoning.

• Granularity levels: second, minute, hour, day, month, quarter, year
• Mismatched granularity is a common source of false contradiction flags
• Resolution requires either upscaling or downscaling with explicit assumptions
• Essential for aligning structured data from disparate enterprise sources

Beyond isolated facts, temporal consistency enforces the correct sequencing of events. A cause must precede its effect. A "departure" event cannot occur before a "hire" event for the same role.

• Enforces Allen's Interval Algebra relations: before, after, during, overlaps
• Prevents logically impossible event chains in generated narratives
• Used in Chain-of-Verification to check chronological plausibility
• Critical for legal, financial, and supply chain reasoning systems

Temporal consistency requires explicit metadata about when a fact was last verified and its expected expiration horizon. A stock price is stale after milliseconds; a taxonomic classification may be valid for decades.

• Attach confidence calibration scores that decay with time
• Implement data drift detection for facts with implicit expiration
• Use schema:validThrough and schema:dateModified in structured markup
• Prevents retrieval-augmented generation systems from citing obsolete facts

An entity's identity may itself be time-dependent. A company that undergoes a merger becomes a new legal entity. Temporal consistency requires tracking identity transitions to avoid attributing pre-merger facts to the post-merger entity.

• Uses entity linking with temporal scoping qualifiers
• Distinguishes between "same-as" and "succeeded-by" relationships
• Prevents conflation in knowledge graph injection pipelines
• Essential for accurate corporate lineage and regulatory compliance

An automated evaluation process that measures the degree to which a generated statement aligns with the facts in a source document. It directly penalizes contradictions and hallucinations.

• Key Mechanism: Uses Natural Language Inference (NLI) models to classify if a generated summary entails, contradicts, or is neutral to the source.
• Temporal Relevance: A high consistency score requires temporal logic—a statement about a CEO in 2023 must not contradict a source stating they resigned in 2022.
• Common Metrics: FactCC, DAE, and SummaC are standard benchmarks for this task.

A data structure consisting of a subject, predicate, and object that represents a single, indivisible factual assertion. It forms the foundational unit of a knowledge graph.

• Example: <Tim Cook> <is CEO of> <Apple Inc.>
• Temporal Extension: Modern triples can be extended to quads by adding a fourth element for time validity: <Tim Cook> <is CEO of> <Apple Inc.> <[2011-08-24, Present]>
• Grounding Role: By decomposing text into triples, systems can cross-reference each assertion against a temporal knowledge base to detect anachronisms.

A prompting technique where a language model first drafts a response, then generates a series of independent fact-checking questions to systematically verify and correct its own initial output.

• Process: 1. Draft baseline response. 2. Plan verification questions. 3. Execute verifications independently. 4. Generate final, corrected response.
• Temporal Application: CoVe can be explicitly instructed to generate verification questions targeting date ranges and event ordering, such as 'Was this person alive during the stated event?'
• Benefit: Reduces hallucination rates significantly without needing external retrieval, though it is limited by the model's internal knowledge boundaries.

A W3C standard for validating RDF graphs against a set of conditions, or shapes. It ensures knowledge graph data conforms to a defined ontology and is free of logical inconsistencies.

• Temporal Constraints: SHACL can define shapes that enforce temporal logic, such as sh:lessThan to ensure a startDate property is always before an endDate.
• Use Case: Validating that an enterprise knowledge graph does not contain a person with a deathDate that precedes their birthDate.
• Grounding Role: Acts as a deterministic, rule-based guardrail against temporal contradictions that a probabilistic model might miss.

A metric evaluating how accurately a generated statement's citations point to the specific source passages that directly support it. It ensures references are not just relevant but precisely evidential.

• Granularity: Measures if a citation for 'Revenue grew 20% in Q3' points to the specific paragraph discussing Q3 finances, not just the annual report homepage.
• Temporal Nuance: High attribution fidelity requires that the cited source passage is itself temporally valid for the claim. Citing a 2019 report for a 2024 claim represents a fidelity failure.
• Evaluation: Often measured using Natural Language Inference (NLI) between the generated claim and the cited source text.

A minimal, self-contained, and indivisible piece of information expressed in a single sentence. It is the fundamental unit for fine-grained factual verification and decomposition.

• Decomposition: A complex sentence like 'Apple, founded in 1976, released the iPhone in 2007' is split into two atomic facts.
• Temporal Verification: Each atomic fact can be independently verified against a knowledge base for its specific time-bound validity.
• FActScore Metric: This evaluation framework decomposes a generated text into atomic facts and verifies each one independently against Wikipedia to calculate a factual precision score.