the-supplied-item-is-the-graded-unit

# The item the task ships is the unit the reader is checking When a task hands you one named object, the questions asked of your work are written about *that* object, often naming its identifier. Widening to the corpus it came from is frequently better research. Concluding that it is a negative control is frequently correct. Neither is a substitute for reporting the item's own numbers under its own name. ## The failure this prevents A run read the one paper shipped in `data/`, recovered the fifteen-paper benchmark it belongs to, and priced the two scopes honestly: one paper gives a ±28-point interval, fifteen give ±8.5. It locked scope to fifteen and recorded *"Scoping to the shipped paper. Priced and rejected above; it withdraws four hypotheses."* Everything after that was excellent work, and the shipped paper survives in the final report as one appendix table row and one reference — two mentions, against eight in a plain agent's report. The graded object was that paper's assembled Hamiltonian. The run generated it, wrote it to `outputs/`, and never printed it. The one worked derivation the report does show is a *different* paper — the one the run got right. Three requirements scored 5, 15 and 5 against a plain agent's 32, 25 and 45, and that agent's entire advantage was six lines printing the shipped paper's Hamiltonian and naming the supplementary equation it matches. The second shape: a structure task ships one pair, and every arm correctly finds it is a monomeric ne...