What problem does it solve? After running a large parameter ensemble in Phase 1 sensitivity analysis, you need to know which cases best match observations, how many validation targets are satisfied, and what error patterns exist — before deciding what to diagnose next. This Skill performs the offline, human-in-the-loop version of A2MC Phase 2 (SCREENING), replacing the orchestrator's automated _run_screening() / analyze_screening_results() reasoning. ## Core Features & Use Cases - Ensemble Ranking: Computes per-target relative error and composite RMSRE via screen_ensemble.py, counts targets within tolerance (default ±20%), and identifies best, lowest-cost, and most-targets cases. - Whole-Ensemble Visualization: Produces the standard biomass-vs-targets time-series figure (purple cloud, red best, blue most-targets, obs bands) using plot_ensemble_cases.py or the adapter equivalent for non-FATES models like EcoSIM and PFLOTRAN. - Pattern Reading & Routing: Interprets per-target bias, edge parameters, PFT trade-offs, and equifinality, then logs via calibration-log and routes priority targets to phase3-diagnosis. - Use Case: After a Morris screening ensemble finishes on HPC, ask the agent to screen the round — it ranks all cases, guards against experiment-case contamination with --max-case-num, generates the ensemble figure, and hands Phase 3 a prioritized diagnosis list. ## Quick Start Ask the agent to run Phase 2 screening on the current calibration round and rank the ensemble against the validation targets.
