A new arXiv preprint introduces FDE-Bench, a benchmark for evaluating end-to-end delivery from underspecified real-world business requests. The authors observe that frontline practitioners often struggle to articulate rules they know implicitly, leaving key requirements unstated in their requests.

Because of this, forward-deployed engineers must actively elicit those requirements before building a solution. FDE-Bench appears designed to test whether AI systems can handle that same ambiguity and deliver a complete result, rather than simply following a precise specification.

The paper is a single source, so there are no conflicting findings to compare. Its significance lies in shifting evaluation toward realistic, messy inputs where the real challenge is figuring out what the user actually needs.