Baseline Intercepts Versus Persona Slopes - Data and Analysis Package

Archival release of the study:

Wiencek, M. (2026). Baseline Intercepts Versus Persona Slopes: Stimulus and Administration Fidelity of Polish Narrative-Biography Personas in Large Language Models. (manuscript revision 2026-07-17)

Supersedes v1.0.0: documentation cleanup (neutral wording throughout, DOI badge and citation metadata added, checksums regenerated). Data, stimuli, and analysis code are unchanged; all released tables reproduce bit-for-bit (CI-gated).

Contents

• Scored data (v20): 1,156 wave-tagged model runs (22-item battery) + 123 runs of the 57-vignette extended battery; human sanity check (N = 7) in aggregate form only, per participant consent.
• Stimuli: 30 Polish narrative biographies with author-declared target profiles (researcher-only YAML headers, verified never to reach any model-facing prompt) + the full TCTM-22/57 vignette source with author keys (stimuli/tctm54.ts).
• Collection pipeline actually used, audit manifest (run_manifest.csv, 1,265 calls with SHA-256 of every prompt and response), prompt-hygiene tests (zero target leakage across 2,884 archived prompts).
• Reproduction: one command regenerates all 40 tables (fixed seed; pinned environment); CI gates: rerun, output diff, SHA-256 checksums.

Raw per-run artifacts (~119 MB) are deposited as a separate restricted-access Zenodo record (third-party instrument copyright).

Licensing: code MIT | author materials CC BY 4.0 | third-party instruments excluded (THIRD_PARTY_NOTICES.md).

Concept DOI: 10.5281/zenodo.21406224