The half-foot that was not there

Every elevation Mathyra extracted from one corpus was slightly wrong, in the same direction, by roughly the same amount. Not noise — a bias. The kind of error that survives review precisely because it is consistent: nothing looks anomalous when everything is shifted together.

The first honest number

A human read 59 points off the sheets by hand, blind to what the software said. Scored against that reading, the engine sat low by 0.44 ft. A sign test on the residuals put the probability of that being chance at roughly one in ten trillion. Something structural was wrong.

Where the error actually lived

An elevation ladder prints numbers beside a grid. To turn a page position into an elevation, you fit a line through those labels. The question nobody had asked out loud: which part of a printed label is the elevation?

The extractor used the centre of each label’s bounding box, because that is what a bounding box gives you. The drafter meant the text baseline — the invisible line the digits sit on, which is where the gridline is drawn. For the typeface in question the baseline sits at 78.3% of the box height, not 50%. The gap between those two conventions, multiplied by the plot scale, is about four tenths of a foot.

The software was reading the drawing correctly. It was reading the wrong part of the letter.

Four routes, one number

A single measurement of a bias is a story. Four independent measurements agreeing is a finding. The correction was estimated from the drawn gridlines, from a shadow run of the whole corpus, from the corrected engine itself, and from an entirely separate instrument that compares extracted profiles against printed vertical-curve callouts — a route that had never seen any of the datum work.

Printed profile callouts

−0.4018 ft

Drawn gridline lattice

−0.4248 ft

Blind human reading

−0.4409 ft

Spread across all three

0.039 ft

After the correction, the profile instrument’s bias fell from −0.4018 ft to −0.0516 ft — measured against printed numbers on sheets, not against ourselves.

What we take from it

Two things. First, a plan set states its own answers often enough that you can check an extractor against the drawing rather than against opinion — printed callouts, gridlines, ruler labels. Second, the errors that matter are rarely loud. This one lived in a default, and it took an outside reading to make it visible at all. That is why the blind reading of fresh points remains the gate on any accuracy claim we make.