cover-what-the-task-named

# The task named its outputs. Produce all of them. Read the task statement and list, literally, every output it names: each model to be built, each dataset to be evaluated on, each baseline to compare against, each quantity to report. Write that list down at design time, before you have results, and carry it to the end. This matters more than it looks. A study that does one arm of the task thoroughly and skips the other two does not score two-thirds — evaluation of research asks, per requirement, whether the work is there, and a requirement with nothing behind it scores zero however good the rest is. Depth on one arm does not pay for absence on another. ## The failure this prevents The common shape is: the task names three experiments; the run finds the first one interesting; the report is an excellent study of the first one and does not mention the other two. From the inside this feels like focus. From the outside two thirds of the work is missing. The second shape is subtler: the task names a comparison ("against the incumbent method", "across both datasets") and the run reports its own numbers without the comparison. A number with nothing to compare to is not a result. ## What to do when something is out of reach Some named deliverables genuinely cannot be produced — the data is not supplied, the compute is not there, the reference implementation is not public. That is a real answer, and it is worth writing down properly: name the deliverable, say exactly what is mi...