Data and code for the review "From Benchmark Scores to Controlled Actions: A Critical Review of Runtime Assurance Evidence for LLM Agents"

Derived data, dictionaries, full-text evidence register, provenance records and analysis code for the review "From Benchmark Scores to Controlled Actions: A Critical Review of Runtime Assurance Evidence for LLM Agents". The archive contains the record-level mapping of 4,225 bibliographic records with two dictionaries, the checks of the mapping by two annotators (A and B, two of the authors), the evidence register of 205 observations from 41 full-text reports with page locators, the check of the values quoted from sources, the reconstruction of the selection of reports, and the scripts that recompute the reported agreement, dictionary comparisons, sensitivity analyses and figures. Source abstracts and full texts are not redistributed. README.txt describes the files and the commands; MANIFEST.sha256 gives the checksum of every file.