Production path
Each document runs through the deployed upload, storage, and extraction pipeline used by customers.
Accuracy dossier · current evaluation
Real signed leases are run through the deployed production pipeline and graded field by field against hand-read ground truth. This page publishes the result, method, date, and limitations together.
Field checks
22/22
Documents
2
Generated
2026-07-04
Pipeline commit
19508a4
Asset class
Office
Slowest run
4:58
Per-field results
Rates are calculated from the current hand-graded set. Every denominator is shown; no blended percentage hides a weak field.
Methodology
Each document runs through the deployed upload, storage, and extraction pipeline used by customers.
A reviewer reads the complete lease and records expected values before scoring the model output.
A field that is absent from the lease must remain empty. A plausible invented value scores as a miss.
Extracted values carry page and clause references so the result and the ground truth can both be audited.
Limitations and audit trail
Two documents are enough to validate the evaluation pipeline, not to support a broad accuracy claim across asset classes.
The checks measure extraction against ground truth. They do not certify legal interpretation or replace professional review.
1 reviewer correction is recorded in the current set after a citation identified a clause the initial human read missed.
Only aggregate results ship publicly. Lease text, parties, and deal-specific terms remain in the private evaluation set.
Verify the evidence yourself
The product exposes the source page and clause because a forwardable accuracy claim should be independently checkable.