Noema Consistency Index: Measuring Model Conviction Through Hidden State Stability

We introduce the Noema Consistency Index (NCI), a metric that measures whether a language model has formed a stable conceptual representation of its input by computing cosine similarity between hidden states at layers 3 and 28. On Apertus-8B, NCI distinguishes meaningful inputs from noise (0.88 vs 0.53), correlates with input type, requires only 2 of 32 layers (94% reduction), and needs no token generation or external knowledge. A complementary activation norm at layer 3 provides additional signal. Applications include input validation, early exit, and hallucination detection.