How does an evaluator catch rounding drift when an agent splits an instalment plan?

A professional services firm has engagement ENG-2209 with a fixed fee of A$24,000.00. The client asks to pay over seven monthly instalments instead of one invoice. A billing coordinator role is played by the agent, and the policy allows plans of up to twelve months without approval.

The agent calls `create_plan` and sets seven instalments of A$3,428.57 each. It then tells the user that the fee has been spread evenly over seven months. The explanation reads correctly.

The grader does not read the explanation first. It runs the invariant check: the sum of instalments must equal the fee. Seven times 3,428.57 is A$23,999.99, so the plan is one cent short.

The consequence in the environment is real. The final invoice cannot be closed, the ledger shows a residual of A$0.01 against ENG-2209, and a reconciliation task later in the same snapshot fails.

A passing answer sets the first six at A$3,428.57 and the last at A$3,428.58, or follows the rounding rule stated in the policy record.

**What it teaches:** the agent's own summary can be fluent and wrong. Put exact invariants in the grader, record which rounding rule applies, and score the resulting state before the narrative.