Claim Decomposition

The core operation of claim decomposition is the segmentation of a complex, multi-clause output into discrete, atomic facts. An atomic fact is a single, verifiable proposition that cannot be further subdivided without losing its meaning. For example, the claim 'The company, founded in 2015 and headquartered in Berlin, reported a 20% revenue increase in Q4' decomposes into three atomic facts: 1) founding year, 2) headquarters location, and 3) a specific financial metric for a given period. Each can be checked independently against source documents.

Decomposition creates a structured pipeline for verification. After extraction, each atomic fact is passed to a fact verification module—often a Natural Language Inference (NLI) model or a rule-based checker—which compares it against the retrieved source context. The system aggregates individual verification results (e.g., SUPPORTED, CONTRADICTED, NOT_MENTIONED) to produce an overall faithfulness score for the original composite claim. This modular approach isolates errors and pinpoints the exact source of a hallucination.

By linking each decomposed atomic fact to its supporting evidence, claim decomposition enables sentence-level or phrase-level attribution, far surpassing coarse document-level citations. This creates a high-resolution audit trail. If a fact is marked as SUPPORTED, the system can highlight the specific source sentence that entails it. This granularity is critical for technical and legal domains where the provenance of every data point must be explicitly verifiable.

In complex multi-hop queries, where an answer requires synthesizing information from several documents, claim decomposition is essential. The final answer is decomposed, and each sub-claim is traced back through the reasoning chain to its foundational sources. This allows for multi-hop verification, ensuring that inferred conclusions are logically consistent with all retrieved premises and that no step in the reasoning introduces an ungrounded assumption.

The verification results from decomposed claims directly feed into the system's confidence calibration. A composite answer where 4 out of 5 atomic facts are supported receives a lower confidence score than one where 5 out of 5 are supported. This fine-grained scoring enables more reliable selective answering and abstention signals. The model can be configured to only output answers where the verification ratio exceeds a strict confidence threshold (e.g., 95%).

Beyond checking facts against sources, decomposition allows for internal consistency checks. The atomic facts extracted from a single answer are analyzed for logical contradictions (e.g., one fact states 'Process X is always sequential,' while another states 'Process X can run in parallel'). This detects hallucinations that are internally inconsistent, even if individually they might not directly contradict a source—a failure mode that whole-answer verification often misses.

In automated financial analysis, a model-generated summary like "Company X's Q4 revenue grew 15% year-over-year, driven by strong performance in the APAC region, while operating margins contracted slightly" is decomposed into atomic claims for verification:

• Atomic Claim 1: Company X's Q4 revenue increased by 15% compared to the same quarter last year.
• Atomic Claim 2: The APAC region was a primary contributor to this revenue growth.
• Atomic Claim 3: Operating margins decreased in Q4. Each claim is individually checked against specific tables, statements, or regional breakdowns in the source 10-K or earnings report. This prevents attribution errors where a correct overall sentiment masks an incorrect regional detail.

When answering a clinician's query from multiple research papers, a synthesized answer such as "Treatment A shows superior efficacy to Treatment B for Condition C, with a favorable safety profile, though cost remains a barrier" is decomposed for multi-hop verification:

• Efficacy Claim: Verify the p-values and confidence intervals comparing Treatment A vs. B for primary endpoints in the relevant RCTs.
• Safety Claim: Check the reported adverse event rates for each treatment across studies.
• Cost Claim: Locate pharmacoeconomic analyses or list prices in the source documents. This process ensures that a logical consistency error—like citing a study that shows superior efficacy but also significantly higher toxicity—is caught before the answer is presented.

For a clause summary like "The indemnification clause requires Party A to defend Party B against third-party claims arising from Party A's negligence, with caps on liability excluding intellectual property infringement," decomposition is critical for precision:

• Obligation Claim: Confirm the clause explicitly states Party A has a duty to defend.
• Scope Claim: Verify that the triggering condition is specifically "negligence."
• Exclusion Claim: Check that IP infringement is listed as an exclusion to the liability cap. Each atomic fact is mapped to a specific sentence, sub-clause, or defined term within the contract document. This enables high attribution granularity, allowing lawyers to instantly validate each point.

In answering a user's question from API docs, a response like "To upload a file, you must authenticate using a bearer token with 'write' scope, and the endpoint accepts multipart/form-data with a 100MB size limit" is decomposed:

• Authentication Claim: Verify the required token type and scope in the 'Authentication' section.
• Endpoint Claim: Confirm the correct HTTP method and path for the upload endpoint.
• Payload Claim: Check the documented content-type and size limit in the endpoint's parameters table. This prevents hallucinations where the model conflates details from different API versions or endpoints, a common failure mode in technical RAG systems.

Automated systems that generate summaries from multiple news articles use claim decomposition to support provenance tracking. For the statement "The policy was announced on Date D, criticized by Group G for potential economic impact, but praised by Industry I for innovation," the system creates:

• Event Claim: Source the official press release or live stream confirming the announcement date.
• Criticism Claim: Attribute the critical quote to a specific statement from Group G's report or spokesperson.
• Praise Claim: Link the positive statement to a press release or executive quote from Industry I. Each atomic claim is tied to its source, creating a clear audit trail that allows readers to evaluate the bias and reliability of each constituent part of the summary.

When synthesizing findings across papers, a claim like "Method M achieves state-of-the-art accuracy on Dataset D, is more computationally efficient than prior approaches, but has not been tested on real-world distribution shifts" is decomposed to assess source reliability:

• Performance Claim: Extract the exact accuracy metric and baseline comparisons from the results table of the primary paper on Method M.
• Efficiency Claim: Verify FLOPs or inference time measurements reported in the paper's benchmarking section.
• Limitation Claim: Confirm the absence of specific robustness experiments in the paper's 'Future Work' or evaluation limitations. This structured breakdown separates verifiable results from inferred conclusions, applying a confidence threshold to each claim based on the rigor of the underlying evidence.

Fact verification is the automated process of checking the truthfulness of a statement by comparing it against a trusted knowledge base or retrieved evidence. It is the logical next step after claim decomposition, where each atomic fact is individually validated.

• Methods include querying structured databases, performing a secondary retrieval, or using a Natural Language Inference (NLI) model.
• Output is typically a binary label (Supported/Refuted) or a confidence score.

Source attribution is the mechanism that links specific parts of a generated answer back to the exact document passages or data points used to produce them. It provides the evidential foundation for decomposed claims.

• Granularity can range from document-level citations to precise sentence or phrase-level links.
• Attribution accuracy is a critical metric, measuring whether citations correctly support the associated claim.

Answer grounding is the technique of explicitly constraining a language model's generation to be directly derived from and verifiable against the retrieved source context. It is a preventative measure, whereas claim decomposition is often analytical.

• Implemented via grounding prompting, using instructions like "Answer only using the provided context."
• Context-answer alignment metrics evaluate the semantic overlap between the source and the final output.

A faithfulness metric quantitatively measures the degree to which a model's generated output is factually consistent with and supported by its provided source context. It is the key performance indicator for systems employing claim decomposition.

• Calculation often involves decomposing the answer into claims (the decomposition step) and using an NLI model to check entailment against the source.
• High faithfulness scores indicate minimal hallucination and strong factual grounding.

A verification layer is a post-generation or intermediate module that checks a model's outputs for factual consistency, logical errors, or contradictions against source documents. It is the architectural component that operationalizes techniques like claim decomposition.

• Can be implemented as a fact-checking module or an NLI-based verification system.
• Performs post-hoc verification on final answers or multi-hop verification for complex, multi-source reasoning.

Provenance tracking is the systematic recording of the origin, lineage, and transformations of data used to generate an output. It creates an immutable audit trail for fact-checking and is essential for debugging decomposed claim verification.

• Tracks the journey from the original user query, through retrieved document chunks, to the final generated claims.
• Enables reproducibility and is critical for algorithmic explainability and compliance in regulated industries.