Skip to content

ACCURACY

What we measure, and what we refuse to guess.

Anyone can say their software reads drawings. The question a contractor should ask is measured against what, and what happens on the drawings it cannot read. Both answers are on this page.

98.1%

FIELD ACCURACY

Across every field on nine drawings with hand-written ground truth.

0

ROOMS INVENTED

Nothing reported that is not on the drawing. This is the number that matters most.

1

ROOM MISSED

Out of the full set. We publish this beside the others.

What the number means

A benchmark is only worth the ground truth behind it. Ours is nine real drawings whose every room name, every area, and every project metric was transcribed by hand before the software was ever pointed at them. The benchmark then compares field by field, never page by page or document by document, and reports the percentage that match.

It runs on every change. A change that improves one drawing and quietly breaks another shows up as a lower number before it ships, which is the entire point of having one.

Two drawings can both be "read successfully" and one of them be useless. Field accuracy is the only measure that tells them apart.

What this number does not tell you

  • Nine drawings is a small corpus. It is large enough to catch a regression before it ships, which is what it is for, and too small to promise you a percentage on your own set. Treat it as a floor we hold ourselves to.
  • “Zero rooms invented” is a measurement, not a guarantee. It is zero across this corpus on the current build. The engine is designed to refuse before it fills in, which is why the figure holds, but no one can promise it will be zero on a drawing nobody has seen.
  • It measures reading, and never judgement. A field read correctly off a drawing that is itself wrong will be scored as a hit. The benchmark tells you whether the software saw what is on the sheet; it cannot tell you whether the sheet is right.
  • Code findings are not in it. This measures extracted fields — room names, areas, project metrics. Whether a code rule was applied correctly is a separate question, and every pack’s depth is published on Regions & rate packs.
  • Your own set is the only number that matters. The free trial runs the same benchmark logic against your drawings, which is a better answer than ours.

Why "invented" is the number to watch

A missed room costs you a few minutes. An invented one costs you a phone call to an architect about a room that does not exist, and once that happens, every other line in the report has to be checked by hand, which is the whole saving gone. So the engine is built to refuse instead of filling in. Three examples of it refusing:

A SETBACK IT WILL NOT READ

On a zoning statistics block, a number sitting beside "required front yard setback" can be the setback or it can be a dimension the drawing happens to place at that height. Where the row is ambiguous the field is left empty. Reading it wrong once produced a 0.17 m front yard — an instant zoning failure, invented, on a compliant house.

INSULATION IT WILL NOT FAIL YOU ON

The code sets a minimum effective thermal resistance. A drawing note usually gives the nominal rating of the insulation alone, and the two differ by about a third. Comparing them directly would report a correctly designed wall as a critical code failure, so a nominal value is reported for review with its arithmetic shown.

A DOOR RULE IT WILL NOT MISAPPLY

Code minimums for door width apply to a particular door — the required exit, the barrier-free entrance, and no further. Applied indiscriminately, an ordinary house produced dozens of critical failures. Where the schedule does not say which door is which, the check is skipped.

What this number does not cover

Nine drawings is nine drawings. It is enough to catch a change that breaks something, and it is not a claim about every drawing ever produced. The benchmark set is residential and light-commercial work in Ontario and British Columbia; a drawing unlike those is not represented in the figure.

Reading a drawing and knowing which code it answers to are separate questions, and the second one has its own honest answer. Eight code packs ship today, across Canada, the United States and England. Ontario’s is the deepest, at 105 rules; the younger packs run fewer checks, and the report says what kind of pack you are reading before it shows you a finding, so a pass never reads as more than it is. Rate and cost data runs wider than the rules do — 43 cities across Canada, the US and Europe — but a code pack for anywhere else in Europe is authored on request. The full coverage table →

The number covers what is extracted — rooms, areas, dimensions, schedules, project metrics. It is not a claim that every code rule is automated. Some rules genuinely cannot be settled from a drawing: an air-tightness test result, an energy label at occupancy, a solar-ready roof provision. Those are shown as review items and the trade selector says so before you run the audit, instead of presenting a checklist as if it were a finding.

An audit is a second set of eyes, not a stamp. It does not replace the professional who seals the drawings.

What it is tested against

1,500+

REAL DRAWINGS TESTED

Permit sets, tender sets, architectural, structural and electrical — in English, Spanish and Portuguese. Not synthetic files.

14

DRAWING FORMATS READ

Including the DWG the architect actually sent, instead of a flattened PDF export.

230/230

CAD FILES READ WITHOUT ERROR

Every CAD drawing in the corpus, including six that older tooling silently truncated to nothing.

3,000+

AUTOMATED TESTS

Each one written against a specific wrong answer the software produced on a real drawing.

What a test grades — the article on every row is the part that gets checked
The only benchmark that settles it

Run it on your own drawings.

The only benchmark that settles it for you is your own set. Fourteen days, and nothing is charged until they end.

Start free 14-day trial → See the Blueprint Auditor