human-in-the-loop gating

This study addresses the inefficiency of manual auditing in corporate governance and the challenges in validating automated alternatives by proposing a task replacement system within a Digital Governance Framework (DGF). Methodologically, it introduces a residual workload threshold as the criterion for task substitution, establishing an information sufficiency gating mechanism. Architecturally, the framework integrates intelligent agents, rule engines, and evidence services, automating auditing workflows through forward deployment engineering and constructing the DGF-Bench benchmark for multi-model evaluation. Experimental results demonstrate that models such as Gemini achieve success rates up to 94.98% under strict gating conditions, confirming the technical feasibility of automating specific governance tasks and offering an effective paradigm for enterprise digital governance.

This work addresses the risks of generative interfaces in high-stakes settings—such as hallucinations, semantic distortion, bias, and accessibility barriers—that arise from insufficient human oversight and undermine user understanding and control. To mitigate these issues, the paper proposes a “supervision-by-design” architecture that deeply integrates human judgment into the generative pipeline. This framework employs automated risk detection across dimensions including readability, semantic fidelity, factual consistency, and accessibility compliance, coupled with explicit UI controls and a tiered escalation protocol that triggers mandatory human review upon violations. By combining human-in-the-loop (HITL) interventions with human-on-the-loop (HOTL) continuous monitoring, the approach establishes a scalable, verifiable governance loop. The resulting system significantly enhances transparency, reliability, and inclusivity, thereby strengthening accountability and user agency in high-risk human-AI collaborative decision-making.

This work addresses the absence of a formal specification language that clearly delineates responsibility boundaries, approval checkpoints, and governance constraints between humans and AI agents throughout the software development lifecycle (SDLC). To this end, we propose a domain-specific protocol language tailored for AI-augmented SDLCs, which decouples policy intent from mechanistic implementation through a formal grammar, well-formedness conditions, operational semantics, and execution invariants to mitigate collaborative uncertainty. Our approach innovatively formalizes the principle of separation of duties as a 2+N team model and natively integrates Kleene closure with protocol self-consistency verification. Theoretical analysis demonstrates that structured execution reduces system failure rates to the weighted product of agent and verifier failure probabilities. A prototype implementation confirms the feasibility and effectiveness of the proposed method.

This study addresses the accountability deficit in agent development arising from the misalignment between platform controls and service provider terms. By analyzing four categories of tools and policy documents, we map workflow responsibilities and propose a novel grid model distinguishing verification mandates from executors. This framework reveals structural deficiencies in approval mechanisms, demonstrating that responsibility gaps have evolved from human oversight to inherent product attributes. Empirical findings indicate conflicting accountabilities across layers, contradictory attribution logic, and insufficient efficacy of approval artifacts. To support further research, we release a comprehensive dataset and validation scripts as open-source resources. Collectively, this work provides both theoretical grounding and empirical evidence necessary for reconstructing accountability frameworks in agent-based software systems, highlighting the urgent need to address systemic rather than incidental failures in current governance architectures.

Enterprise-scale general-purpose agents lack built-in, reusable governance mechanisms for autonomous cross-tool operation, making it difficult to satisfy requirements for compliance, auditability, and behavioral controllability. This work proposes the CUGA policy system, which embeds runtime governance capabilities into five critical checkpoints of the agent execution pipeline—intent protection, playbook guidance, tool invocation control, human approval gating, and output formatting—through a modular “policy-as-code” architecture. Without requiring model fine-tuning, CUGA enables proactive, continuous, and structured behavior control. By integrating typed governance primitives, dynamic playbook injection, and human-in-the-loop approval, the system effectively blocks malicious requests, enforces structured tool sequences, and triggers manual review for high-risk operations in healthcare scenarios, significantly enhancing policy adherence, execution consistency, and deployment safety.

本文针对工具使用型AI代理的授权架构问题，通过提出七个结构要求和四层参考架构来解决授权决策点、执行及问责机制不足的问题。

This study addresses the security blind spot in coding agents where approval logs fail to cover transitive side effects, proposing a closed-loop mechanism that binds approvals to workflow effect boundaries. Methodologically, it introduces the first formalized closed-loop approval security analysis framework, defining "approval laundering" as a quantifiable failure model and incorporating a source-supported effect prediction freezing strategy prior to authorization. The system implementation integrates information-theoretic limit derivations with pre-tool-call techniques. Experimental results demonstrate that the proposed approach significantly reduces residual unlogged events, achieving an effect prediction macro-recall of 0.926. By effectively curbing unrecorded persistent side effects, this work provides verifiable security guarantees for agent systems.

This study addresses the deficiency in Business Process Management (BPM) education, where overlooking AI’s central design role leaves students ill-equipped for process-level AI decision-making. To bridge this gap, we propose the “AI Decision Checkpoint” framework, which treats AI as a first-class citizen and explicitly distinguishes task-level automation from process-level value creation. By integrating large language models, retrieval-augmented generation, intelligent agents, and process modeling and mining techniques, the framework establishes a six-module closed-loop pedagogical system. Furthermore, it embeds AI evaluation and decision-making stages into a customer onboarding case study. Preliminary validation demonstrates that this approach effectively enhances students’ clarity in AI-integrated decision-making and their comprehension of process-level value creation.

本文提出Agile-V Assurance Spine，通过权威源配置文件、工件绑定、风险适当独立性和时效性等方法解决工程生命周期中对代理输出的正当行动问题。

研究针对Loopjacking问题，通过分析和测试多种代理产品，提出确保审批操作与执行操作一致性的方法，以防止未授权状态变更或表示不匹配。