Accuracy benchmark

Last run: July 2026 · Reproducible from our eval harness

Most "AI document extraction" accuracy claims are unverifiable. Ours isn't: we test against rent rolls filed with the U.S. Securities and Exchange Commission — public documents anyone can check — and publish exactly what we measured.

6/6

documents passed

95/95

fields exactly correct

100%

field accuracy

Methodology

Corpus: 6 real rent rolls and rent-roll summary tables from SEC EDGAR filings (8-K loan exhibits and CMBS prospectus supplements) — office, multifamily, student housing, and mixed portfolios.
Goldens: expected values were hand-transcribed from each source document by a human, independently of the extractor.
Scoring: every golden field (unit numbers, tenants, square footage, lease dates, rents, deposits, stated totals) must match the extraction exactly — no partial credit, no fuzzy matching on numbers.
Verification layer: beyond extraction, each document passes our deterministic checks — recomputed totals reconciled against stated totals, unit counts cross-checked.

Per-document results

Document	Type	Source filing	Result
Hartman portfolio rent roll (2018)	Multi-property tenant-level rent roll, 8-K loan exhibit	SEC 0001446687-18-000051	PASS
Uptown Tower rent roll (2013)	Office unit/tenant rent roll, 8-K loan schedule	SEC 0000928953-16-000005	PASS
Woolworth Building tenants	Tenant-level table from underwritten rent roll, CMBS prospectus	SEC 0000950136-05-003719	PASS
Mazza GrandMarc unit mix	Student housing (232 units) unit-mix summary, CMBS prospectus	SEC 0001539497-13-001242	PASS
Luxe at 1548 unit mix	Multifamily (54 units) unit-mix summary, CMBS prospectus	SEC 0001539497-15-001777	PASS
Granada Gardens unit mix	Multifamily (940 units) unit-mix summary, CMBS prospectus	SEC 0001539497-14-001559	PASS

What this does and doesn't claim

This benchmark shows exact-match performance on real, messy, public rent-roll documents — including OCR-degraded source text (one document literally spells "LIFESTYLES" as "LJFESTYLES"; we extract what's printed, faithfully). It does not claim every possible document will extract perfectly. That's why every extraction ships with a verification report: recomputed totals reconciled against stated totals, and any value the model can't read with confidence is flagged, never guessed. You always know which fields to double-check.

The corpus grows over time. Want your (anonymized) document format represented? Send it over.

Try it on your own rent roll — 3 free