Geocoding 40,000 well records without losing our minds

Every orphan well in our queue needs a dot on the map. That sounds trivial until you see the source material: a county plat with a hand-drawn X, a legal description in section-township-range, and a spreadsheet latitude with the sign flipped. All three describe the same wellhead, and all three disagree by up to half a mile.

We reconcile about 40,000 of these records. This is how the pipeline works.

The three location languages

Well locations arrive in exactly three dialects, and every record speaks at least one badly:

• PLSS descriptions ("NW 1/4 of Section 14, T3N R7W") — precise in theory, but quarter-quarter calls are routinely wrong by one subdivision.
• Decimal coordinates — convenient and frequently sign-flipped, datum-shifted, or transcribed with swapped digits.
• Plat references ("see Map Book 4, Page 11") — accurate but trapped in scanned PDFs.

Our rule: no single source is ever trusted alone. A location only graduates to the map when two independent sources agree within tolerance.

The reconciliation ladder

Each record climbs a ladder of evidence, cheapest first:

1. Parse the PLSS description into a centroid with an uncertainty radius derived from the subdivision depth.
2. Project any decimal coordinates and check the datum; NAD27 stragglers get shifted before comparison.
3. Compare PLSS centroid against coordinates. Agreement within the uncertainty radius plus 100 meters graduates the record.
4. Escalate disagreements to plat lookup, which is expensive and human-assisted, so only about 6% of records reach this rung.

The ladder matters because it bounds cost. Full plat research on every record would take years; applied to 6%, it takes one analyst about a week per county batch.

What we do with disagreement

Disagreements cluster beautifully. When we plot them, three patterns cover nearly everything:

Pattern	Share	Fix
Sign-flipped longitude	41%	Auto-correct, flag for review
Off-by-one subdivision	33%	Expand uncertainty, keep both
Datum shift (NAD27)	18%	Reproject, auto-accept
Genuinely ambiguous	8%	Plat escalation

The table is the whole game. Once we classified errors instead of treating each as a snowflake, auto-resolution jumped from 52% to 94%.

Uncertainty is a feature

Every dot on our map carries its uncertainty radius, and we show it. Contractors planning site visits consistently tell us the honesty is the point: a 400-meter circle they can trust beats a false-precision pin every time. Precision you cannot defend is just decoration.