Project Ariadne: A Structural Causal Framework for Auditing Faithfulness in LLM Agents

As Large Language Model (LLM) agents are increasingly tasked with high-stakes autonomous decision-making, the transparency of their reasoning processes has become a critical safety concern. While $\mathit{\text{Chain-of-Thought}}$ (CoT) prompting allows agents to generate human-readable reasoning traces, it remains unclear whether these traces are $\mathbf{\text{faithful}}$ generative drivers of the model’s output or merely $\mathbf{\text{post-hoc}}\mathbf{\text{rationalizations}}$. We introduce $\mathbf{\text{Project}}\mathbf{\text{Ariadne}}$, a novel XAI framework that utilizes Structural Causal Models (SCMs) and counterfactual logic to audit the causal integrity of agentic reasoning. Unlike existing interpretability methods that rely on surface-level textual similarity, Project Ariadne performs $\mathbf{\text{hard}}\mathbf{\text{interventions}}$ ($do$-calculus) on intermediate reasoning nodes – systematically inverting logic, negating premises, and reversing factual claims – to measure the $\mathbf{\text{Causal}}\mathbf{\text{Sensitivity}}$ ($\phi$) of the terminal answer. Our empirical evaluation of state-of-the-art models reveals a persistent $\mathit{\text{Faithfulness}}\mathit{\text{Gap}}$. We define and detect a widespread failure mode termed $\mathbf{\text{Causal}}\mathbf{\text{Decoupling}}$, where agents exhibit a violation density ($\rho$) of up to $0.77$ in factual and scientific domains. In these instances, agents arrive at identical conclusions despite contradictory internal logic, proving that their reasoning traces function as “Reasoning Theater” while decision-making is governed by latent parametric priors. Our findings suggest that current agentic architectures are inherently prone to unfaithful explanation, and we propose the Ariadne Score as a new benchmark for aligning stated logic with model action.