TALAMANA · THE AI LITERACY MAP FOR ARCHITECTURE AND DESIGN · Studio Practice · AGE 19—21 · FACTUAL · EVOLVING
Quantities and schedules
No count from a vision model enters the bill until a person has recounted it.
When to use
Any bill of quantities, door or window schedule, finishes schedule or area statement: any table where a number becomes a rupee.
The method
One: counts and areas come from the building model's own schedule export, or from measured drawings by a person. A vision model's count of a PDF can start the list; it never prices it — no count from a vision model enters the bill until a person has recounted it. Two: use the language model for what it is good at — structuring the table, checking it for internal inconsistencies, flagging a door that appears in the schedule and not on the plan, formatting the output. Three: every figure traces to a source you can name: a schedule export, a measured sheet, a rate from a document. Four: spot-check — pick five items and recount them by hand; if one is wrong, recount all. Five: lock the words that carry responsibility, as the protect-the-text card sets out, and let the model move only the layout.
Watch for this
A fluent bill with plausible numbers. Plausible is what these models produce, and plausible is not what a contractor prices. One wrong count in a schedule means a site instruction, a variation, and an argument about who pays.
Try it
Give a vision model a floor plan and ask it to count the doors, windows and rooms. Count them yourself. Compare. Then ask yourself what that error rate would cost across a forty-sheet set.
Prove it
Show one bill-of-quantities extract in which every line names its source, and point to the five items you recounted.
How it works
Counting on a drawing is exactly what the most recent benchmark built from real architectural plans tests — doors, windows, bedrooms, toilets, across a hundred and twenty floor plans — and the benchmark exists because the counts could not be assumed. A structured building model does not have this problem: its schedule is a query over objects, not a reading of a picture, which is why step one routes the number there. EVOLVING because drawing-reading models are being trained and re-tested on every release; the method does not depend on which way the scores move.
What this idea builds on
What this idea opens up
Sources
Open this idea on the map · The complete map · Logika · RBDS AI Lab, India · revised every edition.