Title & Document Intelligence
Extraction is the commodity layer — four vendors sell it credibly. What matters here is what happens after: every field carries a page-and-span citation, anything under the confidence floor routes to a person, and no agent is permitted to render a title opinion.
Filtered by the active basin.
Under 0.75 confidence — held for a human, never auto-committed.
Pre-1950 cursive is the hardest class; error rates run several times printed text.
Title opinions are a legal act. The mesh assembles, counsel signs.
Each one carries a page-and-span citation.
Below the 0.75 floor.
Sign-off is blocked while 4 fields sit below the confidence floor. Verify them first — an attorney should not be asked to sign over a guess.
Reading faster does not clear the bottleneck
The constraint on a title chain is not transcription speed — it is the judgment about what the instruments together mean, and who is willing to sign that.
A field without a citation is unusable
If a landman cannot jump to the page and span an assertion came from, the assertion has to be re-derived by hand. That erases the saving.
The floor is a design choice, not a limitation
Auto-committing a 0.6-confidence bearing is worse than leaving it blank. A blank is visibly missing; a wrong bearing propagates silently into acreage.