We introduce NLDD — a metric for causal faithfulness in chain-of-thought. We identify a Reasoning Horizon k* at 70–85% of chain length, beyond which CoT becomes post-hoc rationalization rather than genuine reasoning.

We formalize the divergence between gradient attribution and causal importance. Hidden Heroes are causally critical but gradient-invisible. Gradient Bloats are prominent in gradients but causally inert. Standard methods miss up to 70% of important components.

K-means clustering of 56,000+ NYC Marathon finishers reveals 5 pacing archetypes — from even-pacers (25.9%) to severe crashers (2.7%). The crisis zone: runners who fade 20%+ from starting pace.

Comparative study: linear regression vs. BERT for political bias detection. HDBSCAN + SimCSE embeddings reveal cluster geometry that linear models fail to capture — with meaningful partisan separation in the embedding space.