Field notes
What we found
while building.
Notes from making drawings machine-readable: the errors that hid in defaults, the arguments for refusing to guess, and the measurements that changed our minds. Written for engineers, with the numbers attached.
Which governs when the plan sheets and the 3-D model disagree
Six agency documents, read in their own words, on whether the plan sheets or the 3-D design model governs when the two disagree at a station.
Read the note →We could not find an accuracy number for plan-derived models — and ours would be wrong too
We could not find a published accuracy figure for PDF plan sets turned into 3D surfaces. Why scoring against the designer's model measures the wrong thing.
Read the note →We read 434 pages of plan sets and got 332 cross-sections
We ran 434 plan-set pages through our reader and digitized 332 stationed cross-sections. Here is the full count, including the set that gave up 28 lines.
Read the note →How to check a digitized cross-section against the sheet it came from
A five-step protocol for verifying digitized cross-sections against the plan set: freeze the points, read them blind by hand, score residuals, run a sign test.
Read the note →One comma
A regex that could not read a comma threw away an elevation printed 6,550. It took one set from zero cross-sections to 26 — which still refuse to ship.
Read the note →A style is not a surface
On one corridor one line style draws up to seven lines per cross section; on another, eight to fifteen, and none of 111 sections is unambiguous.
Read the note →We found a fix that reached 76 more sections and did not merge it
A wider band window reached 162 sections, but the 76 new ones score median 1.106 ft against 0.323 ft on the same set. We did not merge it, and here is why.
Read the note →Everything on the sheet that is not the drawing
Our extractor scored gridlines as ground and the gate reported 7 to 9 ft errors. Before we stripped it out, 47% of one Oregon export was grid. What that cost.
Read the note →The contradiction was ours
On one set, 305 refusals for the same reason: two elevations at one offset. It looked like the plans arguing with themselves. It was our stroke-joining code.
Read the note →There is no such job title
We scraped 909 heavy civil job postings looking for the person who builds machine control models. Across 596 distinct titles, not one names the job.
Read the note →The half-foot that was not there
A systematic elevation error hid inside a rounding decision nobody made on purpose. Finding it took a sign test; explaining it took a font metric.
Read the note →Refusing to guess
Most extraction software is judged on how much it produces. We think the more useful number is how much it declines to produce, and why.
Read the note →The same station on four sheets
Plan, profile, cross-section, control. One physical place, four drawings — and the first time we compared them, they disagreed by sixteen feet.
Read the note →