Privilege-Asymmetric Retrieval for LLM-Based Document Identification: corpus manifest, task definitions, pre-registrations, run logs, boundary-probe and fault-injection sources, and analysis scripts

Companion data release for the manuscript "Privilege-Asymmetric Retrieval for LLM-Based Document Identification: Separating Search Utility from Source-File Access" (submitted to the Journal of Information Security and Applications, 2026).The archive contains everything needed to reproduce every number in the manuscript and its supplementary material: the corpus manifest (URL and SHA-256 for each of the 2,059 public korea.kr document attachments; the documents themselves are public government releases and are not redistributed), the 25 pre-registered known-item task definitions with their pre-run hash, the pre-registration documents with dated amendments, the agent harness and condition definitions, every raw run record (including provider-error rows and the post-hoc runs labelled _fixed, _limit3, _diag and _confined), the analysis scripts with their outputs, and the boundary-probe and fault-injection sources with entitlement dumps. See README.md inside the archive for the file map.Not included: source documents, SQLite index files, compiled probe binaries. The developer's account name is replaced by "developer" in paths.