What does a shortcut in a task pack look like, and how would we find it?

An evaluation lead is testing an agent on utility connection requests. Each case, such as CASE-1180, must be routed to one of five queues: `residential-standard`, `residential-network-extension`, `commercial-small`, `commercial-large` and `engineering-review`.

The agent routes nearly every case correctly. The lead is suspicious because its event logs show few policy checks. She runs an ablation: the same tasks with `submitted_via` removed, then with `raw_text` removed.

Without `submitted_via`, performance drops sharply. In the generated sample, requests from the installer portal were almost always `commercial-small`, and web-form requests were almost always `residential-standard`. The channel predicted the queue, so the agent did not need to read the address, load or supply type.

The lead asks for a revised sample in which channel is independent of queue, and adds cases where an installer-portal request is residential and a web-form request is commercial. She also requests answer keys that list the required checks, so a correct route reached without the connection-load check is scored as a process failure.

**What it teaches**

- Strong outcome results with thin process traces deserve scrutiny. - Run ablations on incidental fields such as channel, sender and formatting. - Ask for distributions of fields against labels as part of sample review. - Answer keys that record required checks catch right answers reached for the wrong reasons.