What problem does it solve? Once you have observed a behavior in a large language model, it is hard to decide how to investigate its internal cause. This Skill provides a macro-level strategy layer that selects and orders research directions for explaining the mechanism behind a validated phenomenon. ## Core Features & Use Cases - Six Research Directions: Choose among Location, Causal Intervention, Tuning & Editing, Formation Tracing, Unit Interpretation, and Decision Auditing, each answering a distinct question about the model. - Strategy Composition: Chain directions into complete strategies such as mechanistic evidence (Location + Causal Intervention), capability editing, complete accounts, model explanation, or decision reliability. - Outcome-Aware Planning: When prior investigation records exist, propose only untried or inconclusive directions, avoiding re-testing confirmed or refuted claims. - Use Case: After discovering that a model exhibits a specific behavior, use this Skill to decide whether to first localize the responsible component, then causally validate it via ablation or activation patching, producing a defensible mechanism claim. ## Quick Start Ask the assistant to design mechanism-research strategies for a validated model behavior, specifying any prior investigation outcomes or pinned directions.
