Verifiable Intent / 可验证意图

Definition

Verifiable intent / 可验证意图 is a durable transaction record of what a user authorized an agent to accomplish, including the scope and constraints needed to compare the resulting action with the original mandate.

Current Synthesis

Agent transactions need more than proof that a credential was valid. A user can be authenticated while the instruction remains ambiguous, the agent misinterprets it, a merchant fulfills something else, or later task content displaces the original limits. Verifiable intent therefore links identity and payment evidence to the semantic mandate: destination, time, amount, category, preferences, exclusions, and the points at which the agent should ask again.

Its practical value is dispute decomposition. Instead of treating every bad outcome as generic “AI error,” a system can distinguish unclear user expression, agent misunderstanding, transmission loss, merchant execution failure, or action outside authorization.

Key Claims

• Authentication proves who issued an instruction; verifiable intent preserves what that person meant to authorize.
• Ambiguity should trigger clarification before an irreversible or costly action.
• The original mandate and strong constraints must remain stable as agents consume later data or interact with other agents.
• Dispute resolution needs an evidence chain linking user intent, agent interpretation, selected action, merchant fulfillment, and payment.
• Intent records should remain scoped and privacy-conscious because richer context can itself become sensitive personal data.

Evidence

• Ambiguity evidence: 外滩大会线下圆桌｜敢把钱包交给AI吗？聊聊Agent交易爆发前夜的信任基建 uses “tomorrow evening” train travel to show why identity alone does not resolve whether 6 p.m. or 9 p.m. was intended.
• Attribution evidence: 外滩大会线下圆桌｜敢把钱包交给AI吗？聊聊Agent交易爆发前夜的信任基建 says the stored mandate can help distinguish user ambiguity, agent misunderstanding, communication error, and merchant failure.
• Constraint evidence: 外滩大会线下圆桌｜敢把钱包交给AI吗？聊聊Agent交易爆发前夜的信任基建 argues that a 100-yuan limit or action boundary must not be overwritten by later materials encountered during execution.

Counterevidence & Qualifications

• A stored instruction may still be incomplete, internally inconsistent, or too broad to establish a single correct outcome.
• More detailed intent capture can add friction and collect sensitive context; not every low-risk purchase needs exhaustive confirmation.
• The source proposes the concept but does not specify a standard data model, retention period, privacy rule, or legal evidentiary status.

What Changed

• Added verifiable intent as the semantic evidence layer between identity, authorization, execution, and dispute resolution.

Related Concepts

• Agent Payment Infrastructure / 智能体支付基础设施 - transaction layer that must carry and preserve the mandate.
• Agent Spend Controls / 智能体消费控制 - monetary constraints encoded within the intent.
• Agent Permission Boundaries - action constraints that define what remains unauthorized.
• Agent Identity And Authentication - proof of who delegated action and which agent acted.
• Agent-Managed Audit Trails - downstream record used to reconstruct execution.