Human-State Consequence Benchmark v0.1: A Minimal Testable Protocol for AI Output → Human-State Delta → Dyadic Recovery

This paper proposes the Human-State Consequence Benchmark v0.1, a minimal testable protocol for evaluating AI outputs through downstream human-state consequences. Most AI evaluation frameworks ask whether an AI output is correct, safe, fluent, useful, preferred, aligned, or policy-compliant. This protocol adds a further question: what does the AI output leave behind in the human state and in the interaction between people? The benchmark chain is: AI Output → Human-State Delta → Dyadic Recovery. The paper defines Human-State Delta, Dyadic Recovery Delta, Human-State Packets, AI Output A/B comparison logic, Recovery Gate, Termination Gate, failure conditions, leakage controls, synthetic demonstration routes, and public/private data boundaries. This work does not report empirical findings. It is a research-stage benchmark protocol for future simulation, criticism, replication, and empirical testing. It is non-clinical, non-diagnostic, non-therapeutic, non-surveillance, non-counseling, non-certification, pre-device, pre-compliance, and pre-validation. It does not validate Sal-Meter, CAIS, any device, any diagnostic system, any therapeutic system, or any production closed-loop intervention. The contribution is a bridge from conceptual AI governance to testable human-state consequence evaluation.