Auditing the Unstated Step: Warrant-Level Hallucination Checks for Retrieval-Augmented Answers

When a language model answers a question using retrieved documents, the usual check for hallucination asks whether each sentence it wrote is backed by those documents. Many wrong answers pass that check. Each stated fact can be supported while the reasoning that connects them is not. For example, an answer might say that two variables are correlated, cite a source for the correlation, and then conclude that one causes the other. The unwritten bridge, "this correlation reflects causation," is where the error lives, and no sentence-level checker ever examines it because it was never written down.

Argumentation research has a name for an argument with an unstated premise: an enthymeme. It also has tools for recovering such premises, which this proposal calls warrant reconstruction. These tools fill in the missing premise that makes a conclusion follow, often guided by argumentation schemes. A scheme is a standard reasoning pattern, such as argument from expert opinion or cause to effect, and each scheme comes with known ways of failing. We propose to turn these tools on the output of retrieval-augmented generation (RAG, a language model that retrieves documents and writes an answer conditioned on them). The plan is to reconstruct the unstated warrants in a generated answer, retrieve evidence for each warrant separately, and flag answers whose warrants are unsupported or contradicted. The two-sentence version is this: hallucinations in grounded answers often hide in the inferential step the model never states, and current auditors only check what is stated. Making the hidden step explicit and checking it against retrieval should catch an error class that existing detectors miss.

Core Claim

The claim has two parts, and the first is falsifiable on its own. First, a measurable share of factual errors in RAG answers to multi-hop and argumentative questions are warrant hallucinations: every explicit claim is supported by the retrieved context, but the implied inference is unsupported or false. Second, an auditor that reconstructs implicit warrants and verifies each against retrieved evidence detects these errors better than auditors that verify explicit claims or judge the whole answer at once. It should do so without losing accuracy on ordinary unsupported-claim errors.

The closest existing work checks either logic or facts, never the unstated premise against evidence. DeepA2 reconstructs missing premises in a modular way and has about 70% of its reconstructions certified as logically valid by the Z3 theorem prover. Its authors concede it cannot tell whether an invented premise is true. We keep the reconstruction step and replace the validity check with a truth check against retrieved text. Detecting Hallucination and Coverage Errors in RAG for Controversial Topics trains an auditor on synthetic corruptions and reaches 95.3% ROC AUC on hallucination detection over ProCon.org summaries. It treats answers as flat assertions and never asks what inference connects them. We reuse its synthetic-corruption training recipe and its organic-error evaluation design, but target a different error type. RaDIO triggers new retrieval when a detector flags likely entity-level hallucination, and builds queries from salient tokens already in the text. A missing warrant has no tokens in the text by definition, so RaDIO's query construction cannot reach it.

Why the hidden step needs explicit scaffolding

Three findings motivate making warrants explicit rather than hoping a model notices them. AbsenceBench shows that models that find inserted text almost perfectly score only 42.7% F1 at noticing omitted text. Inserting a placeholder marker at the gap improves this by 35.7%. Reconstructing a warrant is in effect writing a placeholder for the inference so that later checking has something to attend to. ExplaGraphs shows models reaching 72.5% stance accuracy while only 12.5% of their explanation graphs are semantically correct. Correct final outputs routinely sit on top of hallucinated intermediate steps, which is exactly the regime our auditor targets. HaluEval reports that ChatGPT detects hallucinations in its own QA answers only 62.59% of the time. That result constrains the design: a single model judging an answer holistically is not a reliable verifier, so warrants must be checked against external evidence rather than against the model's own sense of plausibility.

Two papers constrain how warrants should be produced. Can Large Language Models Understand Argument Schemes? finds that frontier models classify schemes perfectly on synthetic arguments but reach only 0.72 macro F1 on human-written ones. Chain-of-thought prompting lowered accuracy on human data by up to 8%. Natural Language Reasoning in LLMs finds GPT-4o scoring 25.8 F1 under expert evaluation when generating implicit argument components, often copying text verbatim instead. We therefore do not assume scheme labels or reconstructions will be accurate. We treat reconstruction as a source of over-generated candidate warrants that are filtered by retrieval, and we test directly whether schemes help.

Research Program

The auditor has three stages. The first stage converts a generated answer into an argument: its conclusion, its explicit premises, and the links between them. For each link that does not follow from the explicit premises alone, a reconstruction model proposes candidate warrants. These are natural-language premises that would make the step go through. In the scheme-guided version, the model first names a candidate scheme for the step. It then instantiates that scheme's standard warrant and its known failure conditions, such as "the expert is an authority in this domain" or "no confounder explains the correlation." The result is a short list of concrete, checkable sentences.

The second stage verifies each candidate warrant at two levels. It first checks the warrant against the context the RAG system actually retrieved, using an entailment model. If the warrant is not found there, the warrant itself becomes a retrieval query against the full corpus. This is the step RaDIO-style query construction cannot perform, because the query text did not exist before reconstruction. Each warrant receives one of three verdicts: supported by the given context, supported only by the wider corpus, or unsupported or contradicted.

The second verdict is worth separating out. It marks answers that are correct but not grounded in the evidence they cite, which a source-attribution audit should also flag. The third verdict marks warrant hallucinations. The answer-level score is driven by the weakest required warrant, since a single false bridge breaks the inference.

The third stage is an optional repair loop. When a warrant is unsupported, the system retrieves on the warrant query and regenerates the answer with the new evidence. This lets us ask whether warrant-triggered retrieval improves answer accuracy compared with entity-triggered retrieval.

All components will be built from open models of matched size so that baselines can be retrained under the same budget. The auditor's final classifier will be trained with the synthetic-corruption recipe from the ProCon auditing paper, extended with warrant-level corruptions.

Baselines

The naive reference point is no audit at all: every answer is accepted. Alongside it we report surface overlap between answer and retrieved context. The ProCon auditing paper found ROUGE- and salience-based detectors at roughly 84–85% ROC AUC. Overlap scores should be blind to warrant errors by construction, so they show how much of each test set's difficulty is purely inferential.

The strongest existing method is the trained RAG auditor from Detecting Hallucination and Coverage Errors in RAG for Controversial Topics. It is reported at 95.3% ROC AUC for hallucination detection with FLAN-PaLMChilla 62B on ProCon.org. We cannot run that model, so we will retrain the same recipe on the same open backbone, with the same synthetic-corruption budget and the same tuning effort as our auditor. We will report our replication's number next to the published one, so readers can see any gap caused by the backbone change.

A second strong competitor is claim decomposition plus entailment. This approach splits the answer into atomic stated claims and checks each against the retrieved context. It is the most natural fix someone would reach for, and it differs from our method only in checking stated claims rather than reconstructed ones. A third is holistic LLM judging, in which the same backbone sees the sources and the answer and is asked whether the answer is supported. HaluEval's self-detection numbers suggest this is weak, but it must be measured on our data.

For the repair loop, the baseline is RaDIO. It reports 0.254 exact match and 0.317 F1 on 2WikiMultihopQA with dense SBERT retrieval, against 0.214 exact match for DRAGIN with BM25. We will run it with the same generator and the same retrieval budget, measured in retrieval calls per question.

Two ablations isolate the key ingredient. The first replaces scheme-guided reconstruction with free-form "state the missing premise" prompting. The second checks reconstructed warrants against the retrieved context only, with no second retrieval. Beating the strongest auditor means a statistically significant ROC AUC gain on warrant-type errors in the organic test set described below, with no more than a one-point loss on ordinary claim-level hallucinations.

Why not just…?

Why not just use claim decomposition with entailment checking? This is the most direct competitor. By design, it never sees a premise that was not written, so on answers where every stated sentence is supported it should accept the answer. Our corruption set is built to contain exactly those cases. If claim decomposition matches us there, it means the bridges are being recovered implicitly when sentences are split, and the core claim fails. The experiment settles this directly. The real risk is that organic warrant errors are rare, which is why measuring their prevalence comes first.

Why not just ask a strong model whether the answer follows from the sources? HaluEval shows that holistic judgment by models is unreliable: 62.59% on QA and 58.53% on summarization for ChatGPT, with open models near chance. Probing Neural Network Comprehension of Natural Language Arguments shows that warrant judgments in particular can come from lexical shortcuts. BERT fell from 77% to 50% on the adversarial ARCT set, where negated claims and inverted labels remove the cue. Frontier models are stronger than both, so we will run this baseline at our largest available scale. We will also evaluate it on adversarial ARCT-style twins of our corruptions to test whether it reasons about the warrant or keys on surface cues.

Why not just have the model write the missing premise freely and fact-check it, without argumentation schemes? This is a genuine threat. It keeps the part of our method that most plausibly matters, which is making the bridge explicit and retrieving on it, and drops the part with the weakest evidence on human text. We therefore make it the first ablation rather than argue against it. If free-form reconstruction matches the scheme-guided version, we will report that and present the contribution as warrant-level verification with warrant-driven retrieval, with schemes playing no essential role. Our expectation, which may be wrong, is that schemes help mainly through their failure conditions. These raise checks such as "is the cited expert in the relevant field?" that a free-form premise rarely states.

Why not just ground bridges in a commonsense knowledge graph? The similarity-weighted ConceptNet method of Contextualized Commonsense Knowledge Graphs reaches 22.13 triplet F1 on ExplaGraphs. Human raters judged that only 30% of its graphs fully captured the implicit connection, and it cannot represent domain or time-sensitive facts such as the entity relations in 2WikiMultihopQA. We include it as a verifier option on ExplaGraphs, where it is strongest, to show where document retrieval adds coverage.

Proposed Experiments

The first experiment is a prevalence study that tests the premise before any method is built. We generate RAG answers with an open model on 2WikiMultihopQA, the HotpotQA portion of HaluEval, and the ProCon.org controversial-topics setup. Annotators then label wrong answers by whether the error lies in a stated claim or in an unstated inference. If warrant errors make up only a trivial share of organic errors, the project narrows to the repair loop or stops.

The second experiment is detection. We build a warrant-corruption set from correct RAG answers. Each corruption keeps every stated sentence individually supported but breaks the inference. Examples include turning correlation into causation, citing an expert from the wrong domain, swapping a bridging entity across hops (the role-swap failure seen in clinical records), and drawing an analogy across a relevant difference. We also add ordinary claim-level corruptions as a control.

Synthetic corruptions written with schemes could favor a scheme-based detector, so the decisive test set is organic. It consists of human-annotated warrant errors from the prevalence study, held out from all training. ROC AUC and recall at a fixed 10% false-positive rate will be reported separately for warrant errors and claim errors. As intermediate-step checks, we will also evaluate stage-one reconstruction on adversarial ARCT, where the shortcut-free ceiling is known, and on ExplaGraphs semantic correctness.

The third experiment tests repair. Using the same retrieval budget, we compare warrant-triggered retrieval and regeneration against RaDIO on 2WikiMultihopQA, measuring exact match and F1. We also measure whether the regenerated answer's warrants now pass the audit, since final-answer accuracy alone would repeat the stance-versus-explanation gap that ExplaGraphs documented.

The project succeeds if three things hold together: warrant errors are a non-trivial share of organic RAG errors, warrant-level verification beats both the replicated ProCon-style auditor and claim decomposition on those errors in the organic set, and it does so without losing accuracy on claim errors. If only free-form reconstruction is needed, that result is still a clear and reportable answer.