What problem does it solve? Open-ended interpretability research often stalls at the first step: deciding which model behavior is actually worth explaining. This Skill guides the discovery half of the research loop by systematically surfacing a new, non-obvious, falsifiable behavioral phenomenon in LLM or multimodal models, along with the data strategy to validate it. ## Core Features & Use Cases - Six Discovery Strategies: Transfer known phenomena into high-stakes domains (chemistry, medicine), borrow findings from human sciences, test cross-modal transfer, reuse prior CS results, probe a phenomenon's boundary conditions or causal origin, and run meta-analysis to distill laws. - Quality Bar for Candidates: Every candidate phenomenon is filtered against five criteria — real, non-obvious, specific, robust, and tractable — and sharpened into a one-sentence falsifiable claim with a data/metric plan and a plausible internal locus. - Deduplication Against Prior Work: When given a record of already-explored phenomena (established, conditional, refuted, or inconclusive), it commits to a genuinely new direction and keeps unpicked ideas as a backlog. - Use Case: A researcher asks "find something surprising about how this model behaves in clinical diagnosis." The Skill brainstorms candidates, commits to one falsifiable phenomenon, and hands it off to /mechanism-explore for mechanistic investigation. ## Quick Start Ask the agent to find a novel, testable behavioral phenomenon in your target model and domain, then hand the chosen candidate to /mechanism-explore for mechanistic analysis.
