Rho Experiment Dataset: Full Data and Code of a Preregistered Same-Origin Failure-Correlation Experiment with Five Model Families (22,500 Calls)

Complete dataset and code of a preregistered paired experiment: five frontier model families (GLM-5.3, DeepSeek-Flash, Qwen3.8-Max, Kimi-K3, GPT-6.1-Sol) on 30 bare tasks × 30 repetitions, dual-anchor design (ten paired arms + five single-seat baselines, 22,500 calls, executed 2026-10-28 to 11-01, 100% complete). Includes: all 30 task materials with MD5-frozen manifest, the full 22,500-row verdict ledger, per-call raw responses with metadata (model/tier/verdict/timestamp/MD5), deterministic no-model judges with two-way self-tests, the execution runner, and the full statistics recomputation script. Reproducibility boundary (honest statement): judging and statistics are fully recomputable and all raw responses are public; re-querying the APIs is not expected to reproduce responses verbatim (server-side nondeterminism and model drift). Preregistered protocol (frozen and published before data collection): see related identifier.