Evidence in Virtual Reality Research: Comparability, Reproducibility, and Geometry-Dependent Redirected Walking

Virtual-reality (VR) evidence is difficult to compare when studies differ in problem definition, intervention, outcome, environment, constraints, measurement, and validation design. We developed an evidence architecture before implementing a reproducible software workflow that tests those comparison obligations in a controlled redirected-walking (RDW) benchmark. The frozen evidence library contains 174 verified records, 173 canonical reports, 171 retrieved and assessed reports, 165 direct eligible VR/XR reports, six supporting/context reports, and two reports not retrieved. We operationalize a seven-dimension comparability model and a six-level reproducibility taxonomy. Among eight pre-specified contradiction families, seven were comparability-limited and one was a controlled regime effect; an independent second coder reproduced every dimension-level and family-level classification (8/8 agreement; Cohen’s κ = 1.00 for each coded field). In the RDW benchmark, Vis.-Poly, S2C, and P2R completed 384/400, 378/400, and 381/400 runs, respectively. Every controller pair reversed the sign of its conditional reset-burden difference across four declared geometries, while completion behavior also changed. Median controller-decision cost was 251.448, 2.022, and 3.938 μs on the common runner. The results support regime-aware, failure-aware, and provenance-preserving comparison rather than context-free rankings; no new RDW controller or live-user outcome is claimed.