Everything on the sheet that is not the drawing

Our own extractor was scoring gridlines as ground. The gate that checks our elevations was reporting errors of 7 to 9 ft, and when we went and looked at which lines were producing them, they were the sheet's grid. Not a ground line. Not a breakline. Nothing a dozer would ever cut to. On the Oregon set, before we stripped the furniture, 47% of what we exported was gridlines, and the per-section export cap was quietly dropping real terrain to make room for them.

That is our defect. No agency drew it wrong. A plan sheet is supposed to have a border and a grid on it; the sheet was doing its job. The software was not.

A plan sheet is mostly not the drawing

Look at any cross-section sheet and count what is on it besides the sections. There is a border. A title block with the project number, the sheet number, the scale note and the seal. A grid behind the sections. Match lines. Revision clouds. A legend. A bar scale. The vertical scale ladder up the left side with its elevation labels, which carry a failure mode of their own — the four-digit pattern that discarded every elevation label above 9,999. Station callouts. On some sets, a north arrow and a strip of standard notes that repeat on every sheet in the volume.

To you, none of that is terrain, and you don't even register it as a decision. To a PDF extractor it is all the same thing as the ground line: a stroke with a width, a color, a dash pattern, and a pair of endpoints. The file does not say "this one is ground." It says "draw a line here."

The legend does not rescue this either. On our corpus, legends do not map line styles to surfaces — they name features, not pens. And the layer names that would tell you which strokes belong to the border live in the PDF's layer metadata rather than printed anywhere on the sheet, so a scan or a flattened plot does not carry them at all.

A vertical line is not a surface

The cheapest test we have is that a ground line is a function of offset: one elevation per station offset, left to right. A grid column, a border edge, and the vertical scale ladder all fail that test immediately, because they go straight up.

On Oregon, 994 lines came out with no horizontal extent at all. Pure verticals. They are excluded, and they are counted — we write down how many and why, rather than letting them disappear into a difference between two totals, the same rule the engine follows when it names a reason instead of handing back a number.

The export cap is where this got expensive. There is a limit on how many lines we export per section, and before we stripped the furniture, the cap was being spent on grid. Real ground lines were falling off the bottom of the list so that vertical grid columns could occupy the slots. Nothing errored. The counts looked plausible. The terrain was just gone.

Everything we throw away gets counted

Across the five sets we have exercised end to end, here is the current census and the exclusions that go with it.

Two things about that table matter more than the totals.

First, 332 is the number of sections we exported, not the number that exist. There is no denominator of "sections on these sheets" behind it, so it is not coverage and it is not recall — the same accounting we walk through in what 434 pages of plan sets actually yielded, denominator stated. FL-002 is the honest illustration: 28 lines out of 1 of its 116 pages. We are not claiming 116 pages of FL-002. We are claiming one page and saying so.

Second, none of the files on disk was produced by the current engine. The five deliveries were written by two earlier builds, and the census above counts what those builds produced — nothing here has been re-run under the engine at HEAD, and when it is, these numbers can move.

"Oregon" here is the location, not the owner — it is a Federal Lands Highway project on the Historic Columbia River Highway, pulled from public records. It is not Oregon DOT's work.

The strip that reached into the title block

The specific bug was in how we find the horizontal strip of the page a section is drawn in. On 3 of 180 strips on the test1 set, the search for the top edge of the first section ran off the top of the page and stopped at the page edge instead — so everything printed above that first section, the title block and the standard notes that repeat on every sheet in the volume, got pulled in with it.

We caught it because of a frozen hand reading. Sixty elevations were read off the sheets by a person, by hand, before the reader that scores them was written, and blind to what the software would say. Fifty-nine of them scored. When the furniture came out of the strip, two of those points moved: P012 by 3.29 ft and P041 by 0.11 ft. Everything else held still.

A 3.29 ft move on one point is the whole argument. That is not a rounding difference or a tolerance question. That is a point that had been reading a piece of the title block.

While we were in there we also corrected the measuring tool itself — it had been misreading closed pavement boxes. Against the same 59 hand readings on the same bytes, that took p95 from 0.892 ft to 0.126 ft, and within half a foot from 45 of 59 to 58 of 59. The canonical artifact for that gate still reads FAIL on disk, and the gate is deliberately outside the v1 roster, so nothing we would put in front of anyone rests on those two numbers — they are a check on our own reading, not a claim about accuracy. We are careful about them in both directions: it was an instrument correction, not an engine improvement — the exported data did not change, only the thing measuring it. And the person recorded to the nearest tenth of a foot, so roughly 0.05 ft of that 0.126 is their own rounding. Nobody read the same point twice, so we have never tested whether they would agree with themselves.

Furniture was hiding a second defect, not standing in for it

The most instructive line we have found runs through the frame. On Oregon page 21, station 40575, a ground line traces out across the offsets, comes back over the same offsets, then drops 64.04 ft straight down at offset 39.83 and terminates on the plot frame. All of that is inside a single exported line.

That happens because the routine that stitches short strokes into a continuous line joins them by endpoint proximity and never asks whether the result is still one elevation per offset. It will happily chain the end of a ground line into a border segment that happens to start nearby. On one style family, 37% of what comes out is multi-valued — two elevations at the same offset. We have not ranked the other families against it.

Splitting every line at its reversals does not fix it. We tried: the count of conflicting-elevation refusals went from 305 to 305. The collisions are between the pieces, not inside them. Removing the furniture made the defect visible; it did not remove it.

And the harder problem sits underneath both. A style triple — pen width, color, dash — is what the drafter's pen was set to, not what the pen drew, which is why a pen setting cannot tell you which line is the dirt. On the 86 Oregon sections where we counted line styles, only 19 have a single distinct style, so on the other 67 there is a choice to make about which line is existing ground and which is finished grade. On FL-003, every one of 111 sections carries between 8 and 15 distinct styles. Zero of 111 are unambiguous. (The style counts and the export census run over slightly different populations, which is why 86 here and 111 for FL-003 do not match the table above.) Cleaning the grid off the page does not tell you which of the eight to fifteen styles still on that section is the one you are going to build to.

Where this leaves the work

Mathyra has never shipped a surface. All 26 deliveries produced so far refuse to export: zero points, zero faces, each carrying the same three named reasons why. That is the correct outcome while the questions above are open, and it is what private development looks like from the inside.

The reason we are writing this one up is that the furniture bug is the most ordinary failure in this whole line of work and the easiest to ship past. It produces plausible-looking output. It fails quietly. It only surfaces when somebody reads elevations off the paper by hand and compares them, which is slow and boring and the only thing that actually caught it. If you are evaluating anyone's digitized cross-sections — ours included — ask what percentage of the exported lines came from the drawing, and ask what happened to the ones that got dropped. Both of those numbers should exist. If they do not, no tier of confidence anyone puts on a plan set — including ours — is worth anything.