VALUE Proposes an Independent Value Check Before AI Agents Act

Summary

VALUE is a proposal to place an independent, verifiable value-checking layer between an AI agent and a consequential action. The task agent would continue to interpret the user's objective, plan, and propose actions, while a separately governed layer would judge whether each action is acceptable under identified value systems before releasing authority. The proposed VALUE Broker would control access to credentials, APIs, tools, or external systems, making a negative decision enforceable rather than merely advisory. Its risk-adaptive design includes Fast review for routine reversible actions, Guard review for higher-impact or uncertain actions, Sentinel monitoring for suspicious trajectories and boundary bypass attempts, and escalation to a person when the evaluator lacks sufficient confidence or authority. For bounded multi-step work, the broker could issue a scoped execution token limited by tools, destinations, time, amounts, or data classes. The proposal does not assume that all organisations share one complete value system. It outlines collective, plural, and federated governance models in which evaluators identify their value systems and policy versions, while relying parties set trusted authorities, assurance levels, and quorum rules. A VALUE Attestation Object would bind a verdict to a canonical action description, context, destination, constraints, and freshness data, with signatures, trust chains, and revocation information supporting verification. Existing attestation standards such as IETF RATS and EAT are presented as possible infrastructure, while C2PA is kept separate as a provenance layer. The article stresses that cryptographic validity proves what was evaluated and by whom, not that the judgement was correct. Missing context, misleading descriptions, compromised evaluators, replayed approvals, and governance capture remain risks. The proposed starting point is a coding-agent pilot in which the broker controls repository writes, secrets, CI changes, and external publication. The pilot would measure evaluator accuracy, calibration, abstention, latency, replay resistance, evasion, verifier availability, and the cost of human escalation before broader standardisation.