We ran a real 21-page apartment plan set through our own tool — here's exactly what happened

Most "case studies" you'll read from a takeoff-software vendor are curated wins. This one isn't curated — it's a real end-to-end run against a real 21-page apartment building plan set (a multi-building project we're calling "Freeport Square" here), uploaded through the live product, results exported, numbers reported exactly as they came out. Including the part that didn't fully resolve.

The headline number

Every wall on the plan set was measured — that part worked cleanly, no errors, no crashes, real processing time, real CSV export. Of the 25 distinct wall types the plan contained, 2 (both wood-stud partition types) came back with a full matched material spec: stud gauge, track, drywall sheet count, insulation bag count, sheet citations and all.

The other 23 wall types didn't get guessed. They got flagged.

This is the part that actually matters more than the headline percentage. The remaining 23 wall types — the majority of the total linear footage — came back with an honest note attached to every single one: "no legend text matched this thickness — confirm manually." No stud gauge was invented. No CMU block size was assumed. The tool measured the geometry correctly and then declined to guess the spec it couldn't confirm.

We didn't stop at the number — we found the actual cause

Rather than publish the first-run number and move on, we opened the plan PDF's own text layer directly to find out why. Here's what we found: this plan set's wall-type legend is printed on the same sheet as the door schedule, with the two tables' columns interleaved in the drawing layout. When text is extracted from that sheet, the door-schedule and wall-type-legend text comes out scrambled together instead of in clean reading order — a structural quirk of how this specific architect laid out that one sheet, not a generic flaw in how plan sets are typically formatted.

What this actually means if you're evaluating a takeoff tool

• A tool that never shows you anything to double-check isn't more accurate — it just gives you no way to notice when it's wrong.
• "68.9%" is not a general accuracy number for this product — it is specific to our Freeport Square test set, which has an unusually laid-out wall-schedule sheet. Other plan sets will differ.
• The real question to ask any takeoff tool isn't "what's your accuracy percentage" — it's "what happens to the readings you can't confirm." Ours get labeled and handed back to you. They don't get silently guessed into a number that looks clean on a spreadsheet.

We're publishing this the same way we published the real bug we found in our own insulation detection — because a specific, inspectable failure mode you can verify is worth more than a marketing number you can't.

Frequently asked questions

What was tested?

A real 21-page apartment plan set (Freeport Square), run end to end through ReadMyBlueprint. This page covers the first run, which exposed a wall-legend text problem on one sheet. In the tour run on the same set, 68.9% of walls matched automatically and 99.8% after one Confirm all.

Why did so few walls match on the first run?

The wall-type legend shares a sheet with the door schedule and the two tables' columns are interleaved, so the extracted text came out scrambled. It is a layout quirk of that one sheet, not a general property of plan sets.

Is this a general accuracy number?

No. It is one plan set. Plans that draw partition types as graphic details instead of text legends can match far less, so expect to confirm walls by hand on those.