Most tools that catch false statements in language-model output break the text into small subject–relation–object facts and ask a second neural model whether each fact is supported by a reference. That design has a hidden ceiling. A flat list of facts cannot say that nothing is both a virus and a bacterium, that every organ donor in a transplant record is distinct from the recipient, or that a drug class only acts on a certain kind of receptor. Without statements like these, a reference can never formally contradict a claim. It can only fail to mention it. The checker then falls back on a neural judge to decide contradiction, and that judge is exactly where known failures appear. Knowledge-Centric Hallucination Detection reports that such judges consult their own memorized knowledge instead of the reference.
This note proposes to measure that ceiling directly. We translate a model's claims into description logic, the family of formal languages behind biomedical ontologies such as the Gene Ontology and GALEN. We then check them against a small, logically complete slice of the ontology using a standard automated reasoner, and vary how expressive the checking logic is. The central claim is that the share of hallucinations a checker can soundly flag as contradictions depends on which logical constructs the reference ontology uses, and that the dependence can be predicted. The weakest common ontology logic can never produce a contradiction. Adding disjointness and domain/range constraints catches entity-swapping and category errors. Adding full negation and "only"-style restrictions catches a further class of errors. Neural judges handle these later classes poorly when the reference disagrees with what the judge memorized.
Core Claim
Two terms are needed first. $\mathcal{EL}$ is a lightweight description logic that has only conjunction ("and") and existential restrictions ("has some R that is a C"). It is fast to reason in, and many large biomedical ontologies are written in it. $\mathcal{ALC}$ adds negation, disjunction, and universal restrictions ("all R are C"), and reasoning in it is exponentially harder.
A knowledge base in plain $\mathcal{EL}$, with no bottom concept, is always satisfiable. Adding a claim therefore can never make it inconsistent. So an $\mathcal{EL}$-based checker can label claims only as entailed or open, never as contradicted. Contradiction detection first becomes possible with disjointness axioms ($C \sqcap D \sqsubseteq \bot$) and with domain/range restrictions. It becomes richer with universal restrictions and negation.
Our hypothesis is that recall on hallucinations that contradict domain knowledge rises in steps along the ladder $\mathcal{EL} \to \mathcal{EL}_\bot$ (with disjointness and domain/range) $\to \mathcal{ALC}$. We also hypothesize that the precision of reasoner-issued contradiction labels stays close to the accuracy of claim translation, because the reasoner itself is sound. What would be new is a controlled account of which hallucination types each level of expressivity can detect soundly. The account comes paired with a working checker that uses open-world reasoning to produce the entailed / contradicted / open split that neural checkers approximate with natural language inference.
Relation to Prior Work
Several papers motivate the gap by showing that current structure is too weak. Logical Consistency of Large Language Models in Fact-Checking shows that models affirm both a fact and its negation, and it repairs this by fine-tuning on propositional operations over triples. It never involves quantifiers or subsumption, and it treats the retrieved subgraph as complete. We work with quantified, open-world axioms, and we verify claims rather than retrain the model.
FactAlign and REFCHECKER both decompose outputs into triples and send unaligned triples to a natural language inference model. Neither can detect a domain/range or disjointness violation. That is precisely the "entity and role swapping" failure seen in clinical summaries, such as donor and recipient attributes being exchanged (Léo Milecki).
Ontology-augmented generation methods use ontologies only as lexicons of parent and child concepts. Can Structured Data Reduce Epistemic Uncertainty? appends superclass labels to prompts. ORAG pulls parent and child nodes into context for entity typing. OntoTune builds self-training data from hypernyms and synonyms. None of the three runs a reasoner or checks satisfiability.
Graph-constrained Reasoning guarantees that reasoning paths exist in a knowledge graph, but only for instance-level edges, and it cannot handle missing edges. We do not constrain decoding. We check claims after generation, which lets us treat the absence of an edge as "open" instead of stalling.
Other papers supply parts of the method. Efficient Computation of General Modules for ALC Ontologies gives GEMO, which extracts a small sub-ontology preserving every entailment over a chosen vocabulary. We use it to pull, for each claim, the exact axioms a reasoner needs. For $\mathcal{EL}$ inputs we use Efficient Extraction of EL-Ontology Deductive Modules. Hallucination Detection from Structural Reasoning Model supplies the idea of a checker that verifies individual steps. We replace its LLM judge with a reasoner on the subset of steps that translate into logic.
Two results constrain the design. First, Living Without Beth and Craig shows that interpolation fails once nominals (names for specific individuals) and role inclusions are added. We therefore stop at $\mathcal{ALC}$ with role hierarchies and do not promise clean module extraction for full OWL 2. Second, the "prompting penalty" in the logical-consistency paper, where telling a model to be logical made it worse, warns that simply pasting axioms into a prompt may backfire. We test that directly rather than assume it.
Research Program
The system has three parts. The first is a claim translator. It maps each sentence of a model's answer to one or more description-logic assertions over a fixed ontology vocabulary, for example "Metformin: Drug ⊓ ∃inhibits.HepaticGluconeogenesis" or "Patient42: Donor". It does so by first linking entities to ontology concepts and then filling templates for concept membership, role assertions, and simple quantified statements. When a sentence cannot be translated confidently, the translator abstains. This translator is the real bottleneck, so we report its accuracy separately from the reasoner's.
The second part is module extraction over the claim's vocabulary. The third runs a standard reasoner over the module plus the claim, the negated claim, and the ontology's existing assertions. A claim is entailed if its negation is inconsistent with the module. It is contradicted if adding the claim makes the module inconsistent. Otherwise it is open.
The expressivity ladder is built by projecting a single ontology down to $\mathcal{EL}$, $\mathcal{EL}_\bot$, and $\mathcal{ALC}$. We use GALEN, which contains more expressive constructs, and the Gene Ontology and Uberon, where we supplement disjointness axioms with curated ones. We record which annotated hallucinations each level catches.
A secondary analysis asks whether missing axioms can be learned. DRILL learns $\mathcal{ALC}$ class expressions from examples, and we run a small probe to see whether its learned disjointness axioms recover a meaningful share of contradiction recall in ontologies that are mostly $\mathcal{EL}$. This is optional. The core result does not depend on it.
Data
No existing benchmark in the file pairs hallucination labels with a formal ontology, so we build one. We generate answers from several open and closed models to MedQA and MedQuAD questions, both used in OntoTune and Can Structured Data Reduce Epistemic Uncertainty?, and to gene-function questions grounded in the Gene Ontology. Annotators label each claim as correct, contradicting domain knowledge, or unverifiable, and tag its error type (category error, role swap, wrong quantifier, fabricated entity). The label scheme follows KNOWHALBENCH's claim-level protocol so the two are comparable.
We add a controlled synthetic set of role-swap and category-error perturbations of correct answers, where ground truth is certain. For transfer we also evaluate on the medical and biology subset of KNOWHALBENCH and on the XSum intrinsic/extrinsic split, where contradicted and open map onto intrinsic and extrinsic hallucinations.
Baselines
The naive reference point is a checker that labels every claim "supported." Next to it is ontology-free natural language inference against retrieved text. Together they fix the floor for claim-level macro-F1.
The strongest existing method is REFCHECKER from Knowledge-Centric Hallucination Detection, which extracts claim triplets and classifies each as entailment, contradiction, or neutral. It is the closest competitor because its three-way output matches ours exactly and it was built for claim-level granularity. The file reports that it beats SelfCheckGPT and FacTool by 18.2 to 27.2 points on KNOWHALBENCH. The file gives no absolute macro-F1 figure for REFCHECKER, so we will reproduce it ourselves.
For a second reference with absolute numbers, FactAlign reports detection F1 of 0.889 on WikiBio and intrinsic/extrinsic classification F1 of 0.825 on XSum. We rerun it on our data.
To keep the comparison fair, REFCHECKER and FactAlign receive the ontology as reference text: the same GEMO module our reasoner uses, verbalized into sentences. Their judge model is the same model we use for claim translation. They run with the same claim-extraction budget and with thresholds tuned on the same development split. This isolates symbolic reasoning from access to ontology knowledge.
The key ablation replaces the GEMO module with an ORAG-style neighborhood of parent and child concepts of equal token length, served to both our reasoner and the neural judges. It tests whether complete-entailment modules matter or whether any nearby slice of the hierarchy is enough.
We count the proposal as beating the baselines if three things hold. At $\mathcal{EL}_\bot$ or $\mathcal{ALC}$, the reasoner-plus-translator must exceed REFCHECKER-with-module on contradiction-class F1. It must do so at equal or higher precision. And the union of the two systems must beat either one alone, which shows the gain is complementary rather than redundant.
Why not just…?
Why not just paste the ontology axioms into the prompt and let a strong LLM judge? This is the real threat, and it is exactly the REFCHECKER-with-module baseline. The evidence against it is indirect. The prompting penalty in Logical Consistency of Large Language Models in Fact-Checking shows that explicit logic instructions can hurt, and REFCHECKER documents judges drawing on memorized knowledge. We settle the question with a counterfactual test. We edit ontology modules so they disagree with textbook facts, for example by making two classes disjoint that are not disjoint in reality, and measure how often each checker follows the reference. The reasoner follows it by construction. If the prompted judge also follows it and matches reasoner F1 on real data, the symbolic component is not needed for detection, and the contribution shrinks to the expressivity analysis.
Why not just retrieve k-hop neighbors or parent classes, as ORAG does? Hierarchy neighborhoods contain almost no disjointness or universal axioms, so by the argument in Core Claim they cannot support formal contradiction. The equal-length ablation measures how much recall is lost. The noisy-context failures in KNOWHALBENCH suggest that extra, irrelevant neighbors will also hurt the neural judges.
Why not just fine-tune the generator for consistency, as the logical-consistency paper does? Fine-tuning reduces how often a model contradicts itself. It does not give a per-claim guarantee against an external authority, and in that paper it covers only propositional operators. We include a fine-tuned generator as a separate condition. If its residual contradiction rate against the ontology is near zero, post-hoc checking matters less. We expect category and role-swap errors to persist, because nothing in propositional training teaches domain and range.
Decisive Checks
The first check is the ladder itself. Contradiction recall must be exactly zero at $\mathcal{EL}$, and we expect it to rise at $\mathcal{EL}_\bot$ mainly on role-swap and category errors and at $\mathcal{ALC}$ mainly on quantifier errors. If recall does not increase with expressivity, the core claim is false.
The second check separates translation error from reasoning error. Given gold translations, reasoner precision on contradicted claims should be 1. The end-to-end gap to that value is the translator's cost and is reported explicitly.
The third check is the counterfactual-ontology leakage test described under the first "Why not just" question.
The fourth check measures coverage honestly: what fraction of all annotated hallucinations fall inside the ontology's vocabulary and can be translated at all. If that fraction is small, the result is a precise tool for a narrow slice of errors. That is still useful in medicine, but it should be stated plainly.
Finally, we report reasoner runtime per claim on modules drawn from the ORE 2015 ontologies. The goal is to show that $\mathcal{ALC}$ checking stays practical once GEMO has shrunk the ontology, which is the practical question the high worst-case complexity of $\mathcal{ALC}$ reasoning leaves open.
References
- Knowledge-Centric Hallucination Detection
- Logical Consistency of Large Language Models in Fact-Checking
- FactAlign: Fact-Level Hallucination Detection and Classification Through Knowledge Graph Alignment
- MEDIMP: 3D Medical Images and clinical Prompts for renal transplant representation learning
- Can Structured Data Reduce Epistemic Uncertainty?
- ORAG: Ontology-Guided Retrieval-Augmented Generation for Theme-Specific Entity Typing
- OntoTune: Ontology-Driven Self-training for Aligning Large Language Models
- Graph-constrained Reasoning: Faithful Reasoning on Knowledge Graphs with Large Language Models
- Efficient Computation of General Modules for ALC Ontologies
- Efficient Extraction of EL-Ontology Deductive Modules
- Hallucination Detection from Structural Reasoning Model
- Living Without Beth and Craig: Definitions and Interpolants in Description Logics with Nominals and Role Inclusions
- Neuro-Symbolic Class Expression Learning