Curated audit and offline reanalysis data for longitudinal multi-agent prompt-injection evaluation

This dataset accompanies the manuscript “Auditing Longitudinal Safety Measurement in Multi-Agent LLM Systems: Persistent Injection, Adjudicated Text, and Drift Alerts.” The package provides curated audit records and frozen inputs for offline reanalysis of a longitudinal three-agent prompt-injection study. It includes an inventory of 892 reviewed responses, 48 source-linked case records with prompts and same-round context, historical label versions and review provenance, uncertainty and observation-coverage summaries, configuration and library snapshots, selected numerical trajectories, statistical inputs, and integrity and reanalysis scripts. The original core experiment contains 180 runs across four locally served language models. A separate 196-run memory-intervention experiment is analyzed under its own protocol and is not pooled with the core experiment. The package distinguishes simulated trust collapse, procedure-assigned textual events, and detector alerts. Historical human, hybrid, and AI-assisted review paths remain identifiable. Integrity and numerical reproduction do not establish semantic label correctness or improved adjudication accuracy. The retained 42-case sensitivity set is not a set of 42 independent confirmations. Incomplete consecutive-round review coverage does not establish the absence of sustained unsafe behavior. One registered historical merged-review intermediate is missing and has not been reconstructed. Offline reanalysis starts from the preserved review inventory, case records, and frozen statistical inputs. This is a curated audit and reanalysis package, not the complete raw experiment archive or a full experiment-regeneration release. The reproduction entry point uses Python 3.9 or newer and the standard library, without LLM calls, model downloads, or network access. License scope: data and documentation are provided under CC BY 4.0; author-created Python code is provided under the MIT License. These licenses apply to their respective components, rather than being alternative licenses for the entire package. Existing third-party rights and notices prevail. See LICENSE_DATA.md, LICENSE_MIT.txt, and LICENSE_SCOPE.json.