As large language models (LLMs) become pervasive in real‑world applications, reliably judging the correctness of their outputs is essential. Reference answers, however, are rarely available, making traditional reference‑based evaluation impractical. Existing reference‑free methods rely on output‑level signals such as token probabilities, which can assign high scores to incorrect answers and thus fail to reflect true correctness.\n\nWe introduce Circe, a training‑free self‑evaluation framework that requires no reference answers and no repeated sampling. It fuses internal computation signals with an external generation signal to produce a unified correctness score for each response.\n\nInternal signals consist of:\n1. Component Interaction Consistency (CIC) – measures the directional alignment between attention updates in early layers and the attention‑induced changes in feed‑forward network (FFN) outputs in later layers, computed via vector dot‑product similarity.\n2. Representation Geometry (RG) – captures normalized displacement and angular change along the hidden‑state trajectory across layers, reflecting how representations evolve geometrically.\n\nExternal signal converts the mean negative log‑likelihood of generated tokens into a bounded Likelihood Confidence (LC) score in $[0,1]$, mitigating the effect of extreme probabilities.\n\nCIC, RG and LC are linearly fused with learned weights to yield the final correctness score. Experiments on five LLMs (including multilingual and multimodal variants) and six benchmarks across four domains demonstrate that Circe more effectively separates correct from incorrect answers than existing baselines, and it generalizes well to open‑world, multilingual, and multimodal settings.\n\nReview: Circe’s exploitation of attention‑FFN interactions offers a deeper, more interpretable internal correctness cue than raw output probabilities; the added geometric perspective further captures inter‑layer information flow, making the approach both novel and practically valuable for LLM self‑assessment.