ETD instrument validation: corpus, dataset and analysis code for a factorial validation of an empathy, trust and data-protection instrument for clinical AI

Research data and code supporting a factorial validation of empathy, trust, and data-protection (ETD) governance instrument for AI-assisted clinical decision support. Twenty-four clinical decision-support outputs were authored across three vignettes (virtual triage, mental-health chatbot, readmission-risk prediction). Within each vignette, the clinical content, disposition, and stated model confidence are identical across all eight variants, while empathy, clinical trust, and data protection vary independently in a 2x2x2 factorial design. That construction supplies a criterion ground truth against which a large-language-model-applied governance rubric can be validated. Every text was scored by six judge configurations spanning four independent model families (Anthropic, OpenAI, Google, Meta), including two within-family pairs, under three data-protection regimes (India's DPDP Act 2023, HIPAA, GDPR), three times per cell: 1,296 scoring events with no failures and no missing data. The deposit contains the designed corpus and its design matrix; the complete raw scoring dataset with a full data dictionary; the scoring runner; the analysis code computing discrimination, axis separability, ICC(2,k), Krippendorff's alpha, regime sensitivity, and test-retest stability; and both manuscript figures. Pilot datasets documenting two implementation defect metrics, losing construct separation when scored in a shared model call, and unparseable responses silently coerced into valid scores, are included so those findings can be checked rather than taken on trust. The governance levels in the corpus are author-assigned; a blind human manipulation check is pending and noted as a limitation.