Temporal Fact Checking

The core function is verifying a fact's truth within a specific temporal context. A claim like "Company X's CEO was Jane Doe" is neither universally true nor false; it must be evaluated against the temporal validity intervals stored in the graph for the ceoOf relationship between Jane Doe and Company X. The check returns a definitive true/false only for the queried point in time or interval.

This process is fundamentally enabled by temporal validity intervals—the [start_time, end_time] metadata attached to facts (triples) in the knowledge graph. These intervals are the source of truth for checks. For example:

• (JaneDoe, ceoOf, CompanyX, [2020-01-15, 2023-11-30])
• (JohnSmith, ceoOf, CompanyX, [2023-12-01, NOW]) A query for the CEO on 2022-05-20 correctly returns Jane Doe; a query for 2024-01-01 returns John Smith.

The system performs automated logical consistency checks against the graph's temporal ontology. It can detect contradictions such as:

• Impossible Simultaneity: Two facts that cannot be true at the same time (e.g., one person being CEO of two competing companies concurrently, if prohibited by the ontology).
• Temporal Violations: Facts that violate defined temporal constraints (e.g., a promotedTo event occurring before a hiredOn date). This is often powered by a temporal reasoning engine applying rules based on formalisms like Allen's Interval Algebra.

High-fidelity temporal fact checking often integrates with an event graph model. Facts are derived from or linked to immutable event records (e.g., PromotionEvent, AcquisitionEvent). Verification can then:

• Trace a factual claim back to its source events for provenance.
• Reconstruct entity state at any past time by replaying relevant events.
• Check the causal consistency of a claim (e.g., a claimed outcome must be temporally after its causative events). This uses the event sourcing pattern.

A key output is temporal provenance—an immutable audit trail explaining why a fact was deemed true or false at a given time. This includes:

• The specific graph triples and their validity intervals used in the evaluation.
• The inference rules or reasoning steps applied.
• The source data or events from which the verifying facts were extracted. This is critical for explainable AI (XAI) and meeting regulatory algorithmic explainability requirements in sectors like finance and healthcare.

The checker must execute time-sensitive queries against a graph that represents evolving truth. This requires specialized query capabilities, such as:

• Temporal SPARQL: Queries using operators like VALID TIME or AT TIME to scope patterns to a specific time.
• As-of Queries: "Show the state of the graph as of 2023-06-01."
• Interval Queries: "Find all facts that were true during Q4 2022." The backend is typically a temporal graph database or temporal graph database as a service (TKGaaS) optimized for these operations.

In regulated industries, verifying the state of an entity at a specific past time is paramount. Temporal fact checking automates compliance audits by querying a TKG to confirm:

• Officer Tenure: Was Person X a Director of Company Y on Date Z?
• Ownership History: What was the percentage ownership of an asset during a merger period?
• Regulatory Status: Was a financial instrument approved for sale at the time of a transaction? This provides an immutable, queryable record for regulators, replacing manual document review.

News claims often involve evolving situations. Temporal fact checking is used to debunk misleading statements by anchoring them to the correct timeline.

• Evolving Statements: Verifying a claim that "The policy was in effect last year" by checking the TKG for the policy's actual temporal validity interval.
• Historical Context: Assessing claims about economic metrics or crime statistics by retrieving the accurate values for the cited time period.
• Event Sequencing: Checking the order of events in a narrative against the event graph to detect fabricated sequences.

Legal cases hinge on establishing facts 'as of' specific dates. Temporal fact checking streamlines e-discovery and evidence validation.

• Contractual Obligations: Determining which version of a contract clause was active when a breach allegedly occurred.
• Digital Asset Provenance: Establishing the chain of ownership for an intellectual property asset at the time of an infringement claim.
• Employment Disputes: Verifying an employee's role, responsibilities, and reporting structure during a specific period relevant to a lawsuit.

Patient records are inherently temporal. Fact checking clinical decisions against a patient's longitudinal TKG improves safety and accuracy.

• Medication Reconciliation: Verifying if a prescribed drug was contraindicated given the patient's active conditions and other medications at that time.
• Treatment Efficacy: Checking if a treatment was administered within the correct protocol window relative to diagnosis.
• Clinical Trial Eligibility: Automatically validating if a patient's health status over a historical interval matches all trial inclusion/exclusion criteria.

Consumers and regulators demand verifiable histories for products. Temporal fact checking validates claims about sustainability and origin.

• Carbon Footprint: Verifying a product's claimed emissions by tracing the temporal provenance of its components and their associated transport/logistics states.
• Ethical Sourcing: Confirming a material was sourced from a certified supplier during the entire production period for a batch.
• Recall Management: Precisely identifying all products manufactured within a faulty component's validity interval, enabling targeted recalls.

Post-incident analysis requires reconstructing the state of a network and its entities at the time of an attack. A TKG serves as a forensic timeline.

• Access Verification: Fact checking if a user account had administrative privileges on a specific server at the time of a data exfiltration event.
• Vulnerability Window: Determining if a known software vulnerability was present and unpatched on a system during the incident window.
• Lateral Movement: Tracing and verifying the sequence of compromised nodes (temporal pattern mining) to understand the attack path.

The core data structure for temporal fact checking. A Temporal Knowledge Graph is a knowledge graph where every fact (a subject-predicate-object triple) is annotated with a temporal validity interval (e.g., (CEO, Apple, 2011-08-24, 2023-08-24)). This explicit time-stamping allows the system to query the state of the world at any specific point in time or over a range, which is essential for verifying time-sensitive claims.

The precise time range during which a specific fact is considered true. It is typically represented as [start_time, end_time], where the end time can be open-ended (e.g., NOW). This interval is the atomic unit of truth in temporal fact checking. A claim like "Tim Cook is CEO of Apple" is verified by checking for a CEO relationship with a validity interval that contains the claim's implied or stated time context.

The primary query language for interrogating a TKG. Temporal SPARQL extends the standard SPARQL query language with temporal operators and functions. Key constructs include:

• VALID TIME clause to specify the time of interest for the query.
• Temporal predicates like BEFORE, DURING, and OVERLAPS to filter facts based on their intervals.
• This allows a fact-checking system to programmatically ask: "What was true on date X?" or "Did relationship Y hold during period Z?"

A system that performs logical inference over temporal data. Beyond simple lookup, a Temporal Reasoning Engine applies rules to derive new time-aware facts or detect inconsistencies. For example, using Allen's Interval Algebra, it can infer that if Event A MEETS Event B and Event B DURING Event C, then Event A BEFORE Event C. This is crucial for checking complex, multi-fact claims that involve sequences or durations.

The machine learning task of inferring missing temporal facts. TKGC models (like Temporal Knowledge Graph Embeddings) learn patterns from existing time-stamped data to predict likely facts for a given time. For fact checking, this can help assess the plausibility of a claim even if no explicit record exists, by evaluating its consistency with learned temporal-relational patterns in the graph.

An alternative TKG model centered on events as first-class entities. Instead of just time-stating static facts, an Event Graph models occurrences (e.g., Election_2020, Merger_Acme_Beta) and links them via temporal (before, after), causal (ledTo), and participative (involvedActor) relationships. This model is particularly powerful for fact checking narrative claims about sequences of occurrences and their causes.