Frozen data underlying “Control Assimilation in Agentic AI”: architecture cases, human-control paths, screening status, and RC reproducibility

Description

This versioned dataset contains the frozen research data supporting the empirical architecture-level evidence synthesis reported in “Control Assimilation in Agentic AI: Formal Human Authority Persists as Operational Access to Control Becomes Technically Mediated.” The deposited material preserves the prespecified first-wave analysis freeze. It includes curated exports of the supplementary-table panels; a status projection of the 6,754 canonical Work IDs and the 1,156-record first-wave P1 cohort; the 31 responsible-human role paths with their data dictionary; the standardized neutral architecture summaries and locked RC decision rule used for independent blinded reclassification; the original and returned RC labels and confidence ratings; the final supplementary material; source/provenance manifests; and checksums. The dataset supports auditability and reproducibility of the frozen architecture coding, screening-state representation, role-path structure, sensitivity outputs, and reproducibility assessment of the locked RC classification rule. The 5,598 P2 records were not substantively adjudicated for this study. The 17-case analytic cohort is a frozen first-wave sample, not a probability sample of deployed agentic AI. The independent blinded reclassification tests reproducibility of the locked RC rule from standardized summaries; it does not constitute independent primary-source extraction, independent taxonomy derivation, external validation, or proof of corpus completeness. No population-prevalence, causal configuration-effect, or longitudinal control-loss inference is supported by this dataset. No third-party full texts, private correspondence, personal contact information, reviewer identities, or internal analyzer workbooks are included. Cited source publications remain subject to their original access and licensing conditions.

Steps to reproduce

1. Use the frozen Master and P1 status projections to reconstruct the reported search-routing and first-wave assessment denominators. Empty or unresolved full-text-screening fields must not be interpreted as substantive exclusions. 2. Use the supplementary-table CSV exports and the 31-role-path dataset with its data dictionary to reproduce the published architecture-level coding and descriptive summaries. 3. Apply the deposited locked RC decision rule to the standardized neutral architecture summaries to reproduce the classification exercise. Compare the returned independent labels and confidence ratings with the frozen original labels. 4. Use the deposited supplementary outputs to verify the reported sensitivity and robustness results. 5. Preserve the prespecified inference boundaries: this dataset does not reopen P2, does not provide independent primary-source extraction or external validation, and does not support population-prevalence, causal-effect, or longitudinal claims.

Institutions

• Universitätsklinikum Tübingen

  Baden-Wurttemberg, Tübingen
• University of Tübingen

  Baden-Wurttemberg, Tübingen
• Otto-von-Guericke-Universität Magdeburg

  Saxony-Anhalt, Magdeburg

Categories

Licence