AcuIndex: A Task-Native Unit and Market Data Architecture for AI Capability

AI models are becoming productive infrastructure, yet the market lacks a task-native unit for measuring verified AI output across models, providers, runtimes, and cost environments. Token prices are not comparable units of productive output; public benchmark scores often omit cost and runtime; private enterprise tests are not publicly comparable. This whitepaper proposes ACU, the AI Capability Unit, as a task-family-specific unit of verified AI productive capacity. ACU translates benchmark-native evidence from language, coding, reasoning, and systems benchmarks into task-family quotes while preserving model-runtime identity, validator type, cost mode, confidence, and data mode. The core economic quantities are ACU per dollar and dollars per ACU. The public layer provides benchmark-derived reference values; the private layer converts customer workloads into posterior productivity estimates and routing policies. AcuIndex v0.2.4 is a public protocol draft and methodology specification. It is not a trading venue, not a financial benchmark, and not a settlement index.