Information, Not Instruction: What Improved an AI Agent Reading Interior Millwork Drawings

Abstract Interior millwork drawings spread each cabinet across elevations, sections and notes, and a priced calculation needs every part with its dimensions. We built an agent that reads such drawings into a structured per-item dossier and leaves every derived quantity to deterministic code, and between July and November 2026 tested more than 250 changes against answer keys fixed before each experiment. This report describes the ones that held, and what they share: each gave the agent new information rather than new instructions. Viewing sheets in tiles raised the share of dimensions found from 15.4% to 96.2% on one sheet, and from 2 to 34 of 62 on another. Using labels from the PDF text layer as fixed anchors raised correct object binding from 48.4% to 93.3%, against 20.6% for placebo labels. One worked example per tool cut malformed tool calls by 73%. A domain fact delivered at the moment a part is booked raised correct cabinet depths from 9 to 16 of 23, and a targeted question from a code detector was followed by a repair in 35% of cases, against 13% for values that were not raised. Generic reading instructions, by contrast, did no better than a length-matched placebo. To our knowledge this is the first measured study of an agent reading interior millwork drawings.