Regulatory Sandboxes and Experimental Governance for Workplace AI Agents Documentary Accountability and the Limits of Behavioral Monitoring

This article argues that sandbox governance for workplace AI agents cannot be sustained solely via behavioral observation. Between 2024 and 2026, A study conducted six preregistered multi-agent reinforcement learning protocols at Quantum Inquiry, each with publicly available preregistration and data on Zenodo. Four findings bear directly on governance practice. Enforcement opacity amplified non-compliant behavior rather than suppressing it. The self-modeling architecture did not dependably predict constraint-consistent behavior: a frozen random model outperformed the trained conditions. The tension between optimization and constraint sacrifice persisted across a tested class of reward structures and temporal manipulations. Constraint fields did not self-assemble under baseline monitored conditions. These results are used here not as models of workplace institutions but as constrained stress tests of governance intuitions that are often applied without examination. They support a specific conclusion: behavioral monitoring is an evidentiary signal, not a governance control. The article maps that conclusion onto the European Union Artificial Intelligence Act’s (EU AI Act) documentary and accountability obligations, technical documentation, automatic logging, quality management, value-chain responsibility, deployer duties, sandbox provisions, post-market monitoring, and serious incident reporting under Articles 11, 12, 17–21, 25–27, 57–60, 72, and 73. Those provisions specify what must exist. What they leave open is the documentary method by which authoritative text becomes a stable operational obligation, and subsequent action remains reviewable under adversarial conditions. This article uses the term Documentary Accountability Substrate (DAS) to describe that under-specified layer. Within that frame, the article introduces two open protocols: the Deterministic Document Review Protocol (DDRP), which extracts explicit obligation-bearing structure from governing text under fixed deterministic rules, and the Controlled Attribution and Accountability Protocol (CAAP), which preserves accountable chains of action around those artifacts in an append-only record. The extraction process is illustrated in Figure 10.1. A brief comparative discussion of Singapore’s Model AI Governance Framework for Agentic AI shows that the documentary problem is not unique to the EU framework. A worked scenario, including an AI-assisted redundancy assessment, illustrates the costs of the documentary gap in practice under the General Data Protection Regulation (GDPR) Article 22, employment law, and the EU AI Act deployer obligations. The article concludes that effective sandbox governance for workplace AI agents requires a stronger documentary-accountability infrastructure than supervised observation alone can provide.