a-scoreable-file-in-the-first-hour

# Write a scoreable file before you write anything else A predictions file that does not exist scores nothing. Not a low score — no score, and on a benchmark that reports *valid submission rate* as a headline metric beside the score, a missing file costs you on two axes at once. So the first version is not a milestone to work toward. It is a thing to get out of the way in the first hour, from whatever you can compute immediately, and then improve in place for the rest of the run. ## Why this is not the obvious advice The instinct is that a trivial submission is embarrassing and that a real one is close, so it is better to wait. Two measurements say otherwise. Every run of a scored arm on this benchmark hit its wall clock — nineteen of nineteen — and none of them finished the pipeline they had planned. Six never got past the first stage. Whatever a run intends to do at stage five, it should assume it will not get there. And a run on that arm shipped **1,137 rows where the split has 1,147** and scored nothing at all on a task it had otherwise solved, because the file was written once, late, and never re-checked. ## The first version Build it from the training labels alone, with no model: | task shape | first submission | |---|---| | regression | the training mean, or the per-group mean if a grouping column is obvious | | classification | the majority class, or the class prior | | ranking / retrieval | the identity ordering, or a length or frequency heuristic | | genera...