Measurement-first, by default
Why we build the eval harness before the product it is meant to grade.
It is easy to ship an AI feature that demos well and quietly fails in the cases that matter. The usual reason is simple: nobody measured the cases that matter before building.
We flip the order. The eval harness comes first. Before Dalang coordinates a single production request, we decide what good looks like, assemble the tasks that represent it, and wire up the scoring. The product is then built to move those numbers.
This changes what a claim means. When we say coordination helps on multi-step work, that statement points at a specific set of tasks and a specific result, not a feeling. If the number is not there, the claim does not ship.
It also decides what we keep private. Which models run inside Dalang, and how they are coordinated, stays internal. We publish results, not routing tables. The evidence is the part that should be public; the plumbing is the part that should be earned.