Reasoning Consistency Scanning: A Framework for Auditing Chain-of-Thought Validity in AI Safety Evaluations
06:00 · July 9, 2026 · arXiv cs.AI RSS
Prior work has shown that chain-of-thought (CoT) reasoning is often unfaithful: a model's stated reasoning does not reliably reflect the process that produced its output. Detecting unfaithfulness, though, requires controlled experimental interventions, which cannot be applied to evaluation transcripts after the fact. We turn instead to a more tractable question that has received less attention: whether the stated reasoning is logically consistent with the answer it accompanies. Unlike faithfulness, consistency can be assessed from a transcript alone, with no intervention. We introduce reasoning consistency scanning, a reusable method for detecting this property in AI safety evaluation transcripts. Our contributions are fourfold. First, we formalize reasoning consistency as distinct from faithfulness and define a six-subtype taxonomy of inconsistency. Second, we build a validated benchmark of 60 transcripts, manually adapted from InstrumentalEval outputs. Third, we implement a working scanner for InspectScout, the first to target this property in safety evaluation transcripts. Fourth, we report results across four generator models and three evaluations from inspect_evals, showing that reasoning inconsistency is present, detectable, and varies systematically across both models and task types.
Summary
Reasoning consistency scanning addresses a practical gap in evaluating large language models during AI safety assessments. While earlier studies have shown that chain-of-thought outputs can be unfaithful to the actual computation that produced an answer, detecting such unfaithfulness typically demands controlled experimental interventions that cannot be applied to existing transcripts. The new framework instead examines whether the stated reasoning steps are logically consistent with the final answer they accompany, a property that can be checked directly from the transcript without further model interaction.
The authors distinguish consistency from faithfulness and introduce a six-subtype taxonomy that classifies the forms of logical mismatch that can appear between reasoning and conclusion. They support the taxonomy with a manually validated benchmark consisting of sixty transcripts derived from InstrumentalEval outputs. A scanner implementing the method has been integrated into InspectScout, allowing automated detection within standard safety evaluation pipelines.
Empirical tests were conducted on outputs from four generator models across three evaluations drawn from the inspect_evals suite. The results indicate that reasoning inconsistencies occur at measurable rates, can be reliably identified by the scanner, and vary systematically both by model and by the type of safety-relevant task. This variability suggests that consistency scanning can serve as a lightweight, post-hoc complement to existing faithfulness checks in AI auditing workflows.
Why it matters
This research is highly relevant for Dutch AI researchers and auditors focusing on AI transparency and safety, aligning with the EU's stringent requirements for trustworthy AI. The proposed framework offers a practical, non-interventional method to evaluate LLM reasoning, which is crucial for developing compliant and reliable AI systems in the Netherlands.
