GroundTruth
The verification layer every Easting product passes through. It is the reason a model is allowed anywhere near a legal document.
Ground truth is the surveyor's and remote-sensing term for verified reference data: the measurement you check everything else against. We use it as the name of the layer that does the checking, because the checking is the product. Extraction is a commodity. Knowing when not to believe the extraction is not.
What it checks
- Traverse closure
- The courses are recomputed from the extracted bearings and distances and walked back to the point of beginning. The gap that remains, divided into the perimeter, is the closure ratio. Curve calls are traversed faithfully rather than approximated away.
- Stated-acreage corroboration
- When a document states an acreage, the computed area has to agree with it. Two independent signals pointing the same way is a much stronger claim than either alone, and a document that closes beautifully while disagreeing with its own acreage is telling you something.
- Geometry sanity
- Self-intersections, degenerate rings, and figures too small to be the thing described get caught before they reach a layer.
- Per-call provenance
- Every course keeps the verbatim text it was read from and a confidence the extractor assigned itself. A reviewer checks the sentence, not the whole document.
What the verdicts mean
- PASS
- The traverse closes within tolerance and the computed area corroborates the acreage the document states.
- REVIEW
- Something disagrees, and the reason is stated: a misclosure, an acreage mismatch, a low-confidence reading. A human decides whether the error is ours or the document's.
- FAIL
- The description cannot be computed as written, for example a course with no bearing. Nothing reaches your map.
The distinction that matters most is inside REVIEW. A traverse that fails to close can mean the extraction misread a digit, or it can mean the 1975 surveyor did. Both are real, and often enough it is the document. GroundTruth does not pretend to know which; it hands a human the misclosure, the suspect courses, and the source text, and gets out of the way. What it will not do is quietly average the difference away.
How it was qualified
The layer was built against a deliberately hostile corpus of recorded deeds: clean modern instruments, 1970s photostats, curve calls in two different notations, a boat slip measured in feet-and-inches, multi-tract conveyances with easement carve-outs. Several documents contain errors in their own original text, including one whose closing bearing disagrees with its opposing course by a full arc-minute.
Across that corpus, on the frontier models we run:
- Every boundary call was extracted correctly, on two independent model families.
- No extraction containing an error ever reached a PASS verdict. That is the number the product lives or dies on, and it is the one we watch.
- Reconstructed parcels were compared against county cadastral geometry by best-fit overlap, with translation and a bounded rotation allowed but no scaling: agreement ran from 0.80 to 1.00 on every parcel we could match.
- Documents with no metes-and-bounds description at all were reported as such rather than hallucinated into geometry.
The same layer sits behind the products still in research. Corridor change detection and site screening produce different geometry from different inputs, and both get the same treatment: recompute, corroborate, and say plainly which outputs a person needs to look at.
The fastest way to judge the validation is to see it run against your own documents.