LLM abstract-screening instrument and complete verdict record for a systematic review of reinforcement-learning controllers deployed on real building equipment

This record contains the abstract-screening instrument and the complete verdict record of the systematic review "When reinforcement learning leaves the simulator: a systematic review of reinforcement-learning controllers deployed on real building equipment" (submitted to Renewable and Sustainable Energy Reviews, 2026). The review screened 6,038 unique Scopus records (5,852 from a keyword search, 186 from the reference list of the nearest field-demonstration review). Titles and abstracts were read by a large language model (Claude Sonnet, Anthropic) against a fixed, versioned rubric; the model's verdicts were triage — the authors reviewed the flagged abstracts and made every final inclusion decision, and every inclusion was decided from the full text. Files: judge_prompt.md (the rubric, byte-exact as served), batch_template.md (the batch format, record entries redacted), judge_verdict_out.schema.json (the enforced verdict schema), and screening_verdicts.zip (one JSON verdict per screened record). Each verdict carries the Scopus ID, decision, evidence tier, a one-sentence reason, a confidence score, and a mandatory verbatim deployment-evidence quotation for every include-tier verdict. No third-party abstracts or full texts are redistributed. See README.md for the reconciliation with the manuscript's PRISMA funnel.