ReflexBench-MW V6 Reviewer Artifact: Matched-World Evaluation of Observer-Participant Decision Consistency

Reviewer artifact for ReflexBench-MW V6, a finite matched-world evaluation of whether language agents preserve decision correctness when their own outputs enter a declared world transition. The release contains seven authored scenario contracts, 14 paired prompts, world-specific tied-optimum action sets, 140 formal model rows from two online model families, 560 machine-semantic sensitivity rows, append-only call provenance, deterministic primary analyses, data and maintenance documentation, and scripts for an author-run CPU reconstruction. The deterministic primary scorer reads no machine-judge or human-rating file. The archive supports reviewer inspection and exact reconstruction of 12 result artifacts. It does not establish different-machine or third-party reproduction, human construct validity, population generalization, deployment safety, or model-family universality.