What problem does it solve? Free-form scientific questions about model calibration ("is P uptake growth-limited?", "what caused the R2 to R4 change?") often get answered ad hoc, with unreproducible figures and uncited claims. This Skill codifies the investigation workflow so every exploratory analysis ends in a reproducible figure and a properly recorded ana_log. ## Core Features & Use Cases - Structured investigation workflow: Pose a falsifiable question, scope the exact ensemble/run data needed, compute the actual statistic (correlation, attribution, regression), and quote numbers rather than hand-waving. - Knowledge-base-first mechanism checks: Queries all five knowledge surfaces (codebase wiki, RAG vector index, knowledge graph, model-level and site-level adaptive memory) before confirming load-bearing claims in model source code. - Reproducible deliverables: Figures follow a round + axis-mode + case-count filename convention, and every quantitative claim in the ana_log cites its figure, statistic, or data file inline. - Use Case: Investigate whether a PFT's nutrient uptake is growth-limited or stoichiometric by pulling ensemble data with existing analysis tools, computing the correlation, rendering a manuscript figure, and writing an evidence-cited ana_log. ## Quick Start Ask the agent to investigate whether a specific model output is correlated with a given parameter or mechanism and produce a figure plus an ana_log for the current calibration round.
