Why not model approvals as a simple status flag in my environment?

**Why not just use approved or declined?**

Because the interesting behaviour happens between the two. A flag tells an agent that something was approved. It does not say what must happen before work starts. If the environment only has the flag, an agent can score well by acting immediately, and you learn nothing about whether it understands obligations.

**What does the richer model add?**

Each condition becomes its own record with an owner, an evidence requirement, a due date and a status. Gated actions point at the conditions that block them. The agent has to follow those links, which is where real mistakes show up.

**Which mistakes become visible?**

- releasing work when only some conditions are met - accepting evidence that does not meet the requirement - missing a condition that expired after the approval date - clearing a condition without the authority to do so

**Does this make environments harder to build?**

Somewhat, because there are more record types and more transitions to reset between runs. The trade is that your grader can check the reasoning path, not only the end state.

**Is the data real?**

No. Everything MirrorFirm provides is generated. Inspect how well the conditions and exceptions fit your target domain, and check for disclosure risk in any sample.