Expertise
Knowing the domain is the hard part.
Any modern model can read text off a page. Knowing what the text means in a 1962 deed, or whether a relationship between two factors is causal or coincidental, is where the work actually is.
What we know about county records
Instruments carry meaning, not just text
A quitclaim deed and a warranty deed transfer very different things. A partial assignment, an affidavit of heirship and a plat dedication each behave differently in a chain of title. Classification is a domain judgement before it's a machine learning problem.
Party roles are structural
Grantor and grantee, lessor and lessee, plaintiff and defendant. Surname-first conventions. Marital status qualifiers that change how an interest passes. Multi-party oil and gas leases where one instrument carries dozens of lessors and the interests sit in an exhibit.
Legal descriptions are a language
Section, township, range and meridian. Quarter-quarter calls. Metes and bounds with bearings and distances. Subdivision, block, lot and plat references. Each is a different grammar, and each has to be parsed into components rather than stored as a string.
Counties don't agree with each other
Formats, abbreviations, export conventions and indexing practices vary by jurisdiction and by era. Handling that is a matter of accumulated, county-specific rules, not a single model that generalises.
Records point to other records
Every deed cites what it derives from — a book and page, a prior instrument number, a plat. A record's meaning depends on the records it references, so extraction has to preserve those links as structured references, not lose them as text.
Historical records are the normal case
Rotated exhibits. Poor scans. Typewritten text over printed forms. Handwritten annotations. The records that matter most for a chain of title are frequently the oldest and worst.
What we know about causal analysis
Correlation is not a mechanism
Two series moving together tells you nothing about whether one acts on the other, or whether a third thing drives both. Most quantitative infrastructure encodes association and leaves the mechanism implicit.
Direction changes the questions you can ask
With directed relationships you can ask what happens if, not only what happened when. That's the difference between reasoning by analogy to past periods and propagating a change through a learned structure.
Explainability has to be structural
Post-hoc attribution tells you what a model did. It doesn't tell you whether the relationship was real. Explanation that arrives after the fact is commentary; explanation built into the structure is evidence.
Confounders are the discipline
The hard part of causal work isn't finding edges, it's ruling out the variable nobody named. Surfacing assumptions and evidence strength for each relationship is what keeps a graph honest.
Common ground
Both domains demand the same three things: extraction that understands what it's reading, structure that survives contact with messy real-world inputs, and reasoning a professional can audit and defend.
That's why one engineering foundation serves both — and why domain knowledge, not model choice, is what separates a system that works from one that produces plausible output.