What problem does it solve? Starting an AutoRL task without grounded evidence leads to hallucinated claims, poor algorithm choices, and wasted effort on unsuitable repositories. This Skill turns an execution-ready task card into a verified evidence package covering relevant papers, benchmarks, and reusable codebases before any planning or coding begins. ## Core Features & Use Cases - Paper Retrieval and Verification: Searches arXiv, Semantic Scholar, Papers with Code, and local notes, verifying exact titles, venues, and URLs while marking unverified claims explicitly. - Codebase Candidate Scoring: Identifies reusable repositories and scores them on environment match, algorithm match, license, install risk, and reproducibility without cloning or installing anything. - Structured Evidence Report: Emits a canonical Markdown evidence report plus optional JSON/JSONL sidecars with coverage scoring, gaps, and handoff recommendations for the strategy decision stage. - Use Case: After clarifying an RL task on a custom Gymnasium environment, run this Skill to produce runs/<task-id>/evidence_report.md listing the top verified papers and safest baseline repositories to inspect. ## Quick Start Use rl-evidence-retrieval to turn my task card at runs/my-task/task_card.md into an evidence report with verified paper and codebase candidates.
