Relevance Is Not Authority: A Sealed Boundary Benchmark of Decision-Confusion in Commercial LLM APIs

Methodology preprint introducing a controlled 1, 000-packet authority-boundary benchmark for commercial LLM decision-model environments. The benchmark separates topical relevance from evidence support, action readiness, publish safety, customer safety, and same-identity authority. The sealed 2026-05-13 run reports 2, 990 scored decisions across OpenAI GPT-5. 5, Anthropic Claude Opus 4. 7, and Cohere Command A Reasoning, with Voyage rerank-2. 5 as the relevance baseline. The reported headline is 214 false-authority decisions before the LatentAtlas guard and 0 after the deterministic boundary guard on this controlled benchmark, with 808/808 expected valid allows preserved. Sealed benchmark manifest: conceptboundaryᵣealₐpi₂0260513. Manifest SHA-256: 06b88b5bf5008f135fe6f361a185efdd58e78f6a9f66d4d308247b86c9a14eb5. This is a methodology preprint, not a peer-reviewed publication and not legal, compliance, or regulatory approval.