Schema Conformance Does Not Guarantee Decision Stability: When the Decision Is Typed but the Interpretation Isn't

Structured outputs constrain an AI system to return a valid category, schema, and numeric scores that downstream software can process, but schema adherence is not the same property as decision stability. This paper reports a controlled supplier-payment classification pilot across three frontier model configurations, using 41 texts and 287 cleaned observations. The corpus comprises one base case, 20 meaning-preserving reformulations, and 20 loaded perturbations. The results distinguish structural validity from reformulation invariance, execution repeatability, and sensitivity to material changes. One tested configuration changed its majority decision on 5 of 20 equivalent reformulations; the three configurations disagreed on 12 of 41 texts; and repeated identical inputs produced a 21.1% pairwise decision-discordance rate in one configuration. The study is a single-domain pilot and does not establish provider-only causal effects, probabilistic calibration, or universal properties of structured-output systems. It concludes that schema conformance and decision stability should be evaluated separately and complemented by deterministic validation and authoritative reference retrieval where operational decisions depend on external facts.