assume-this-stage-is-the-last-one-you-get

# Plan as if the experimentation stage will not happen, because usually it does not The pipeline has a stage for running experiments and a stage for analysing them. On a wall-clocked run they are frequently theoretical. Measured over a nineteen-task scored arm, with a four-hour cap per task: | final stage the run reached | runs | |---|---| | 01 literature survey | 6 | | 02 hypothesis generation | 3 | | 03 study design | 5 | | 04 implementation | 4 | | 06 analysis | 1 | **Nineteen of nineteen hit the cap. None finished the walk.** Six never left the first stage; four of those six spent between 13 and 22 attempts on it. So the predictions file that got scored was, on most tasks, whatever an early stage happened to produce on the way past. That is the environment. Planning as though stage five will arrive is planning for a stage that arrives about one time in twenty. ## What follows **Do the modelling in the stage you are in.** Not because the stage boundaries are wrong, but because a plan whose payoff is two stages away has a low probability of being executed, and a marginally worse model built now has a probability of one. Concretely, at each early stage: - **01, literature survey.** Alongside the sources, run the data loader, print the shapes, and write the trivial baseline submission. The survey is better for it — you now know what the columns actually contain — and if the run ends here, it ends with a valid file. - **02, hypotheses.** Every hypothesis you reg...