What problem does it solve? When a research question concerns how a large language model computes internally — which layer, attention head, neuron, SAE feature, or weight drives a behavior — choosing the right interpretability method is difficult. This Skill routes such questions to the correct method family and enforces a hierarchical loading protocol so analysis is grounded in full method documentation rather than summaries. ## Core Features & Use Cases - Method Family Routing: Eleven families covering vocabulary projection, magnitude analysis, probing, gradient detection, causal attribution, circuit discovery, feature dictionary learning (SAE/transcoder/crosscoder), representation and parameter analysis, SHAP, neural feature learning, and multimodal concept labeling. - Screen-to-Verify Pipelines: Guidance for composing cheap screening methods (magnitude, gradients) with causal verification (patching, ablation, attribution patching) and circuit recovery. - Practical Heuristics: Rules for finding pre-trained SAEs, choosing token positions and intervention layers, and calibrating steering strength. - Use Case: A reviewer asks for mechanistic evidence that a specific attention head drives factual recall. The Skill routes you to Causal Attribution, then to the patching submethod with its demo scripts and paper references. ## Quick Start Ask the agent to identify which internal component of the model is causally responsible for a target behavior and let it route to the appropriate mechanism family.
