Breaking the Perfect Team: An Adversarial Benchmark for Cross-Field Completion Claims

Multi-agent AI systems can satisfy every local contract without establishing that a cross-system business objective occurred or that its completion may be declared. Breaking the Perfect Team (BPT) proposes an adversarial benchmark constitution for assessing this gap without claiming novelty for the well-established failure of local-to-global coherence. BPT distinguishes agent conformance, compositional coherence, independently observed enterprise effects, evidence sufficiency, authority, and final claim release. A Field is treated as an optional mission-relative contract representation, not as an assumed scientific ontology. We specify cross-field episodes, claim-specific oracles, typed receipts, an explicit UNKNOWN state, twelve adversarial failure families with paired controls, and information-matched conventional baselines. An executable reference generator produces 192 deterministic, author-constructed fixtures: 96 positive and 96 negative cases, all with local contracts marked PASS. A local-only rule falsely releases all 96 negatives; a coherence-only rule releases 64; a fully informed conventional verifier and a Field-labelled verifier both hold all 96 while releasing every positive. These engineered results verify fixture consistency, not actual agent performance or independently validated oracle reliability. The paper proposes a blinded, resource-matched evaluation protocol for testing whether claim-explicit composition improves outcomes beyond existing workflow, provenance, and authorization engineering. No Field-specific operational advantage is asserted; parity is an admissible scientific outcome.