Abstract
<title>Abstract</title> <p>Context: Language-model software uses JSON Schema as both an instruction artifact and a validation contract. Software tooling may reorder a Schema without changing validator behavior, yet change model behavior. Objective: We test whether validation-equivalent Schema serializations induce output-distribution changes beyond repeated-call variability, whether the changes exceed a prospectively specified practical screen across deployed systems, and whether provider JSON Mode attenuates them. Method: We derive metamorphic relations from JSON and JSON Schema semantics, issue five repeated calls per representation, and estimate record-level excess disagreement with cluster bootstrap and permutation inference. The core study contains 17,900 successful responses from two gateway aliases and official DeepSeek/Qwen endpoints, a disjoint 200-record decomposition, and a matched 100-record Qwen text/JSON-Mode ablation. Gateway results characterize observed deployments; upstream checkpoint identities were not independently verified. Results: Property-order and additional Schema-member-order effects exceeded 0.05 for the Sonnet gateway (0.058 and 0.077) and Qwen-Plus text deployment (0.126 and 0.123), but not for the GPT gateway (0.033 and 0.038) or DeepSeek (0.013 and 0.018). JSON-Mode effects also exceeded 0.05. Matched JSON-minus-text changes were +0.019 and -0.025; neither was confirmed as a material interaction. Average Schema compliance and leaf-value accuracy did not show universal degradation. Conclusions: Validation equivalence is an insufficient regression oracle for LM-facing Schema artifacts. We release an offline-auditable test workflow and conservative canonicalizer; distributional robustness should be tested separately from syntax and task accuracy.</p>