What's this about?
A benchmark for agents doing real document work — editing, restructuring, cross-referencing office files — with programmatically verifiable success criteria (5↑).
What exactly is "verifiable" scoring?
Checking output correctness by machine rather than by an LLM judge's opinion — a script confirms the document actually has the required properties.
Why is it worth noting?
Documents are where office automation actually lives; verifiable benchmarks are what let RL training loops target the skill directly.