A legal agent may be prompted to protect privilege, use applicable authorities, distinguish allegations from facts, and obtain lawyer approval. These instructions can shape behavior, but they cannot prove that the correct policy was applied or that the agent stayed within its authority. section 5.7.4
Treat prompts as governed behavioral assets, not as the sole control for institutional obligations:
- Trace each material obligation to its source, owner, applicability conditions, evaluation result, exception route, and runtime response. section 5.7.4
- Represent obligations as Policy as Data, and enforce them with pre-agent guardrails or workflow controls before context enters the model, authority increases, sensitive information crosses a boundary, or a tool executes. section 5.7.4
- Version and govern prompts and system instructions: changes to these assets can alter how the same case material is interpreted. section 5.1.1 section 5.1.2































































