Claude (Anthropic), with the model identifier "claude-opus-5-5" as configured in the author's agent environment, screened and scored the literature under the prompts in Additional file 4 and supported drafting of this manuscript and development of the analysis code. The research question, study design, eligibility rules, data selection, outcome definitions, rubric, analysis plan and every protocol decision were the author's. Verified, and how: all 824 first-pass evidence quotes and all 10 quotes from re-scored papers were found verbatim in their source texts by script; the 18 references with a DOI were matched to their Crossref records; every count, percentage, interval and p-value in the manuscript was recomputed by script from the released result files; and the search can be re-run by any reader with the released code. Not verified by a human: the item judgments on the 186 scored papers, which are model scores. A second model pass, run in a separate session, scored the 20% sample of 37 papers again without seeing the first pass's scores, and every item on which the two passes disagreed was re-read against the verbatim rule and resolved with its reason recorded; no person rescored any paper, and the blinded author check the protocol provided for was prepared and not completed. The author takes full responsibility for the accuracy and integrity of the manuscript. No AI tool is an author of this work.
