Perspective independence, not personas, drives LLM teams — and where they reverse

Companion archive for the manuscript of the same title (snapshot v1.0.1). Contains the full analysis pipeline (inference drivers with ER compliance routing: every API call receiving ER-Reason case text is sent through OpenRouter with zero-data-retention routing (provider zdr = true) and data_collection = deny; byte-identical prompt templates including the disclosed prompt-transcription defect and its corrected control; the LLM-judge implementation under the v3 equivalence rules, adjudicating on the official GLM endpoint; statistics and figure scripts), the frozen primary-judge and sensitivity-judge adjudication caches covering every CPC and MedCaseReasoning adjudication reported in the manuscript, and per-case outputs for the CPC and MedCaseReasoning arms (case identifiers one-way hashed; SHA-256 truncated to 16 hex characters). Entries derived from ER-Reason are excluded under the ER-Reason data use agreement; ER-Reason source data are available from PhysioNet under a Credentialed Health Data License.