Agent Approval Laundering: An Approval-to-Action Trace Benchmark and Reproducibility Dataset

This dataset supports the study “Agent Approval Laundering: Transitive Effects Beyond the Approved Invocation.” It examines a structural record-coverage failure in AI coding agents: an approval record may accurately identify the surface command or tool call while omitting the transitive effects activated by the underlying workflow. The dataset contains an append-only ledger of 149 execution attempts and 149 corresponding hash-bound run snapshots, including 139 successful runs.

Its primary benchmark view comprises 111 deduplicated successful agent-task records. It also provides 78 reproducible analytical views covering captured approval objects, decision-time metadata, post-execution traces and artifact evidence, fixed-commit real-project cases, prospective holdout experiments, independent oracle reconstruction, and effect-bound approval evaluation. Raw evidence is kept immutable and separated from regenerated analyses; manifests and SHA-256 digests support integrity checking and reproduction of the study’s principal results.