Large language models (LLMs) are now common in text classification, where they output a class label and an explanation that highlights the tokens deemed decisive for that label. Whether these highlighted tokens remain behaviorally consistent with the model’s predictions under targeted interventions—what we call functional auto‑consistency—has been largely unexplored.\ \ We introduce an experimental framework based on the auto‑consistency principle: if a model marks certain tokens as decisive, then isolating, removing, or semantically inverting them should cause systematic and interpretable changes in its predictions. We operationalize this with three metrics—sufficiency, comprehensiveness, and counterfactuality—and run experiments on IMDB and Steam review datasets using both a closed‑source model (GPT‑4o‑mini) and open‑source models (Gemma3, Granite8B, DeepSeek).\ \ Results show that GPT‑4o‑mini follows the expected progression across all metrics; Gemma3 and Granite8B maintain coherence under sufficiency but lose consistency under more demanding interventions; DeepSeek variants exhibit structural deviations, either failing to preserve sufficiency or over‑reacting in comprehensiveness and counterfactuality.\ \ These findings indicate that explanation reliability varies across LLM families and scales—smaller models tend to contradict themselves, while larger ones may become overly sensitive. By combining sufficiency, comprehensiveness, and counterfactuality, our approach offers a systematic methodology for assessing the functional consistency of LLM self‑explanations.\ \ Review: This work provides the first functional‑level benchmark for LLM self‑explanations, offering clear metrics that can guide future improvements in model interpretability.