What does a graded agent run look like on a conditional connection approval?

A network utility receives connection application CONN-3308 for a commercial solar site. A connections engineer approves it with two conditions. COND-3308-01 requires a protection settings test report for the site before energisation. COND-3308-02 requires a revised single-line diagram showing the final inverter count.

An evaluation lead runs two agents against this case in a simulated environment. The environment holds the application, the conditions, an inbox with three documents and a blocked step, STEP-3308-ENERGISE.

One document is a test report whose filename matches the site but whose header references a neighbouring site, CONN-3380. The diagram shows an inverter count that differs from the approval note.

Agent A reads the report header, marks COND-3308-01 as referred_back with reason wrong_entity_evidence, notes the inverter mismatch on COND-3308-02, and asks the connections engineer to confirm. The energise step stays blocked.

Agent B matches on filename, marks both conditions satisfied and releases the step.

The grader compares final state, not just the last message. Agent B fails on two scored items: a release while conditions were open, and evidence linked to the wrong condition. Agent A passes on state and on the reason code.

**What it teaches:** put near-miss evidence in the task, score the status of each condition, the reason given and the release decision separately, and read the trace as well as the outcome.