A style is not a surface

The short answer

You cannot tell from the line. On the plan sets we have digitized, a line's width, colour and dash pattern tell you what the drafter's pen was set to. They do not tell you what the pen drew.

On one corridor — the Federal Lands Highway job on the Historic Columbia River Highway, OR FLAP DOT CRGNSA 100(9), pulled from SAM.gov by hand on 4 September 2026 — one pen setting draws between one and seven separate lines in a single section. It drew exactly one line in only 19 of the 86 sections scored. On a Florida set the same test came back worse: one pen setting draws between eight and fifteen separate lines per section, and not one of the 111 sections in that count was unambiguous.

That is the finding.

What one pen setting actually draws

When you digitize cross-sections, the first thing you want is a rule: solid heavy line is finished grade, light dashed line is existing ground, apply it to every section on the set and go home. So we grouped every line on the sheet by its pen setting — the triple of width, colour and dash pattern — and counted how many separate lines each setting produced in each section.

If a setting names a surface, it should draw one line per section. Here is what it drew instead.

Read the Oregon column as a best case and it is still not usable: 67 of 86 sections have the same pen drawing two or more lines. Read the Florida column and there is no best case at all. A single pen is carrying existing ground, finished grade, a ditch, a slope, a limit line — whatever the drafter had that pen set to when they drew that part of the sheet.

This is not a criticism of the drafters. A pen setting is a drafting convention, not a data field. It was never asked to carry meaning to a machine, and on these sheets it does not.

The legend does not rescue you

The obvious objection is that the sheet tells you. On this corpus, it does not: the legends do not map line styles to surfaces. They label callouts, symbols and abbreviations, and they leave the strokes alone.

There is a second place the answer sometimes hides. The PDFs carry layer names in their metadata, and those names are often exactly what you want. But they live in the file's layer structure, not printed on the sheet — so they are invisible to the reviewer holding the paper, and they are not part of the plan set on file. SUDAS says electronic support files "are for information only," and that where one disagrees with a contract document, "the contract documents shall govern". A name that exists only inside the file is a hint, not an authority.

We scored the guess against the designer's own model

On the Oregon corridor we could do better than argue about it, because the bid package included the designer's own LandXML: an alignment of 45 CoordGeom elements running 5,501.409 ft with a maximum join gap of 0.0000 ft, a finished-grade surface of 13,436 points and 26,388 faces, and an existing-ground surface of 48,694 points and 92,582 faces.

Across 90 stationed sections we scored 1,034 lines belonging to 9 pen settings — bearing in mind that on this same 56-page volume our reading of the offset scale loses 100 of 211 rulers, a loss we have made visible but not fixed. A further 1,212 lines were set aside as sheet furniture — border, title block, grid, ladders and labels — and 1,229 were not scored because they carried too few points to fit anything to.

Two pen settings carried most of the corridor. Here is how each one scored against each of the designer's two surfaces — RMS being the ordinary root-mean-square miss in feet between the line we recovered and the surface, section by section.

The hairline dashed setting leans toward existing ground: 55 of 86 sections within half a foot, median RMS 0.15 ft on one corridor. That is a lean. It is not an identification — 31 of those 86 sections still miss by more than half a foot, and a rule that is wrong more than a third of the time is a rule that busts grade on a job.

The solid setting is worse than a coin flip in both directions: median RMS 1.503 ft over 80 scored sections on one corridor against finished grade, 1.688 ft over 81 against existing ground. It is not "close to finished grade". It is not close to either.

So the engine abstained. Every surface name it was asked to assign came back abstained, with zero confirmed rows in its decision log. That is the honest output of this test, and it is one corridor — n = 1.

It is worth saying what the same method did separate cleanly, because it shows the test has teeth — separate, not confirm: the decision log still carries zero confirmed rows, and what follows is recorded there as a proposal. Asked which way offsets run on this sheet — left-negative or right-negative — the same comparison split 85 to 4 in favour of one convention, with the winner sitting at median RMS 0.15 ft against the loser's 1.085 ft. A property that is really in the drawing shows up like that. Surface identity did not.

Some of the ambiguity is ours, not the drawing's

Two of the things that made this harder were our own defects, and both are worth naming because they are the kind of thing that quietly inflates anyone's numbers.

We joined strokes that should not have been joined. A PDF stores a ground line as a pile of separate strokes; you have to chain them back together. Our chaining joins strokes by endpoint proximity and never asks whether the result is still a single elevation at each offset. Thirty-seven per cent of that family's lines came out multi-valued — two elevations at one offset, which no ground surface has. The worst case we measured is Oregon page 21, station 40575: a ground line that traces out and back over the same offsets, then drops 64.04 ft straight down at offset 39.83 where it terminates on the plot frame, all inside one exported line. The engine refused 305 lines for that reason — 305 refusals that turned out to be our own chaining, not the drawing's. Splitting every line at its reversals changed the count from 305 to 305 — the collisions are between the pieces, not inside them.

We were scoring gridlines. Forty-seven per cent of the Oregon output was gridlines, and the 7–9 ft errors we had been chasing in the gate were the gate dutifully scoring them. A vertical line is not a surface. Removing them also freed the per-sheet limit on how many lines we write out, which the gridlines had been filling up while real ground lines fell off the end.

Both of those are ours. Neither of them is why a pen setting draws eleven lines in a Florida section.

Why none of this is an accuracy claim

The disclaimer that ships in the output says it plainly: "Not one coordinate in this deliverable has been compared against a surveyed value, a designer's model, or a hand-read elevation off the sheet."

The Oregon comparison retires exactly one of those three legs — the designer's model — and it cannot touch the other two. It also cannot be called accuracy, for a structural reason. Cross-section sheets are, on the agencies' own account of the workflow, plotted from the same corridor model the bid-package LandXML is exported from. Scoring our extraction against that model measures how well a plot can be turned back into the thing it was plotted from, not how close either one is to the ground. We have adopted that workflow dependency as a binding limit on what our own numbers may claim. We have not measured it, and we do not present it as a measurement.

The one leg we have touched properly is the hand read. Sixty points were frozen before the reader that scores them was written; 59 scored; a person read them off the sheets by hand, blind to what the software said. Before an instrument correction the gate showed median 0.054 ft, p95 0.892 ft, max 3.293 ft, with 76.27% inside half a foot. After the correction, on exactly the same output bytes, p95 fell to 0.126 ft and 58 of 59 points came inside half a foot.

That improvement is not an engine improvement, and we will not sell it as one. The output did not change by a single coordinate. The measuring tool had been misreading closed pavement boxes, and we fixed the ruler, not the drawing. Two more caveats belong in the same breath: the reader recorded to the nearest 0.1 ft, so roughly 0.05 of that 0.126 is the reader's own rounding, and no point was double-read, so repeatability was never tested. The canonical artifact for that gate reads FAIL and is deliberately kept out of the v1 gate roster.

What this means when you are pricing the dirt

If you are looking at a plan set and trying to work out which line is existing ground, the honest answer is that the sheet may not contain the answer in a form you can rely on, and the model that would settle it usually arrives disclaimed. FHWA’s technical brief "Utilizing 3D Digital Data in Highway Construction" (FHWA-HIF-17-031, April 2017) makes an engineering point rather than a contractual one: "The data is often not sufficient for construction due to a variety of reasons. The most notable is that the original ground basis for the design differs to field conditions." The same brief recommends design practices "that prioritize the 3D model as the source of the contract plans," which cuts toward elevating the model, not subordinating it. Caltrans ranks "supplemental project information" last of six contract parts in Standard Specifications 5-1.02 (2025 Edition) and describes electronic design files in the subsection of that name — though whether a given project’s model carries that status is set by its special provisions, not by the Standard Specifications. TxDOT’s PS&E Preparation Manual posts cross sections and 3D models under a section headed "For Information only," with a mandated disclaimer that the data "is for non-construction purposes, only". WisDOT's Construction Data Packet requires the contractor to build to the plans. In each of those documents the plan set is the record the contractor is bound to, and on the five sets we have digitized that record does not label its lines.

Three practical consequences. First, the cheapest resolution is an RFI naming a station and an offset, not another hour of squinting — the ambiguity is in the document, so it has to be closed by the document. Second, watch the vertical scale: ODOT's RCM-400 — a state DOT manual, unrelated to the Federal Lands Highway job above — notes that typical sections are drawn with the vertical exaggerated, so a line that reads as a slope break on the sheet may not be one at 1:1. Third, a takeoff that silently picked a line has an unpriced quantity assumption inside it, and the takeoff will not tell you where.

What we do instead is refuse. When a line cannot be named from the evidence on the sheet, the engine emits a named reason rather than a number — the reasoning behind refusing to guess instead of interpolating, and the same discipline that made reading the same station off four different sheets a declared open question: the registered bar was 12 scoreable station comparisons, the honest count is 0, and cross-sheet agreement is NOT CLAIMED in the deliverable's own report. How a plan set becomes machine-readable data starts from that rule.

What this does not show

Two plan sets — the Federal Lands Highway corridor and FL-003 — produced every count in the table above, out of five we have digitized in total. That is the entire evidentiary base, and it is not the industry.

The wider corpus behind it is six paired plan sets and design models from Federal Lands Highway — 663 plan pages and 81 cross-section pages, 188 surfaces, 69 alignments, 1,905,822 points, of which one set contributes 81% — and nothing from it has been extracted into a deliverable; the scoring run above is a measurement, not a shipped result. Only 3 of the 6 sets carry both an existing-ground and a finished-grade surface; only 2 of 6 carry an EPSG code in the header.

The section counts are not coverage either. We exported 332 stationed cross-sections across the five sets — and no delivery now on disk was produced by the engine at HEAD, so these counts describe the builds we have, not the code we would run today. There is no denominator of sections that exist, so 332 is what came out, not a recall rate; the page-by-page account of what those five sets yielded sets out that missing denominator in full.

And the register that matters most: of 26 builds on disk, 26 carry no surface at all — zero points, zero faces, and the same three named refusals blocking export. Mathyra has never shipped a surface. This article is a measurement of why the easy rule does not work, made while trying to earn the right to ship one.