Describe the building.
It measures it, then prices what it measured.

Type and size in, whole build out.

An estimate is built out of assets — buildings and site scopes — each with its own unit mix, floors, spaces, openings and takeoff. The shell is measured once; everything inside is measured per unit type, so a fourplex is not four copies of a duplex.

• Eight building types with real defaults for unit count and storeys
• The IBC construction type is never guessed — unset stays unset, and an unknown value refuses the whole save
• It is snapshotted onto every cost line at creation, so historical costs stay comparable after the asset is edited
• Site work and off-site are their own scopes, with their own earthwork quantities
• Spatial Design lays out levels and placements before an estimate exists — and needs neither a project nor an estimate to do it
• No total is ever stored: gross area, unit count and efficiency are produced on read, so a second number can never disagree with the first

Every line can ride a named measurement. Change the building and it follows.

One registry of measured-quantity kinds is the single list every surface reads — counts, building and unit areas, spaces, openings, trim, takeoff, on-site and off-site earthwork, partitions, fixtures, lighting. Each kind declares which scope owns it, its unit, and the class that decides how it scales across buildings and units.

• A cost line pointing at a measurement moves with the model — unless you typed the number yourself, in which case your number wins
• Every quantity answers one question: is this an assumption or a takeoff? There are deliberately no tiers — a ratio you tuned is still an assumption
• Pure arithmetic over typed geometry is never marked as an assumption. Marking a real measurement as a guess is the worst failure mode here.
• A mixed total is an assumption if any contributing part is on a default — the opposite policy would launder a mostly-assumed total as measured
• Undo remembers what you did, not what the software worked out
• A new quantity kind cannot ship without an explicit provenance decision — the test refuses it

A price is kept with the city, the date, the source and how much you trust it.

Every cost line that graduates out of an estimate or a vendor bid becomes a library observation — a price with the facts that make it comparable later. You start with seeded takeoffs, assemblies, a reference cost database and published rates, and every one of them is built to be replaced by yours.

• Two dates, not one: the date the price was good for, and the date it entered the system
• Six source types, from manual through vendor quote to historical actual, plus a label and notes
• A luxury door never averages with a builder door; a supply-only price never averages with an installed one
• The context is snapshotted at capture — GSF, NSF, units, building count — so the number stays readable years later
• Observations link back to the bid line and invitation they came from, and to the estimate they were first seen in
• AI enrichment fills blanks only and can never overwrite a human edit; the row stays unreviewed until an estimator accepts it

Where one line sits among its real peers.

For a single observation, the median of its like-for-like peer group, and how far above or below the line sits. The segmentation is the whole point: grouping by CSI code would blend unrelated scopes that happen to share a division, so three different electrical lines aren't treated as a market.

• Peers are matched on the canonical item, not the code — plus grade tier, cost composition, asset kind and construction type
• Same unit only. It never forces $/GSF and never averages across units
• An unclassified line gets no number — it refuses rather than guessing
• Only shared rows feed the market; a draft can read it without contributing to it

The work either side of the estimate.