Paraphrastic Resistance: Text Fragility Is Not Decision Instability

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

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 pilot across three frontier models (Anthropic Sonnet 4.5, Google Gemini 3.1 Flash Lite, and OpenAI GPT-5) on a supplier-payment classification task, using a frozen corpus of one base case, 20 meaning-preserving reformulations, and 20 loaded perturbations (287 cleaned observations). Three findings are supported: schema-valid output can be decisionally unstable, as one model changed its majority decision on 5 of 20 equivalent reformulations; loaded changes did not produce the same operational response across vendors, with five perturbations escalated by none; and repeated identical inputs produced different discrete decisions in one model at a 21.1% pairwise discordance rate. The operational conclusion is narrow but consequential: schema validity does not imply decision stability, and these properties should be evaluated separately. Independent work at ICML 2026 provides complementary evidence for the same monitoring problem from a different experimental direction. The study is a single-domain pilot; cross-domain replication is required before paradigm-level claims.