a-deliverable-is-not-an-instruction

# The sheet you were handed is two documents. Only one of them is the task. A task statement that arrives through a harness is rarely only the task. Around the research question sits a second document: how to operate, where to write files, what not to do, what counts as finished. Both are addressed to you, both are written in the imperative, and one of them is not a deliverable. Enumerate the wrong one and the study is planned against phantoms. ## The failure this prevents Measured on a forty-task benchmark arm: the deliverable list drawn off the whole instruction sheet held **337 requirements; the research questions alone held 142**. Fifty-eight per cent of what every run was planning against was the harness talking about itself, and the same five phantoms appeared in all forty runs: > Read & Understand — Study the related work and data to build domain context. > Code & Execute — Implement the analysis, generate figures, and iterate… > Analyze & Report — Interpret the results and produce a publication-quality report. > Your primary goal is to … produce a high-quality `report/report.md`. > Figures are mandatory — generate plots and save to `report/images/`. Every one is a true instruction. Not one is a finding. A plan that answers "Read & Understand" with `figure:2` has spent a figure slot on the fact that reading happened. The damage is not only wasted slots. A plan is checked for a *covering* answer to each listed deliverable, so a list of seventeen forces seventeen ...