Validation Results

Pipeline validation across 16 properties and 8 vendor formats.

At a Glance

Zero extraction errors across 3,732 Units spanning 8 different property management system formats. Every file scored 100/100 on the independent spot check audit.

Test Matrix

Each file was processed through the full pipeline: Upload, Scan, Structure Analysis (Claude Opus), Confirmation, Extraction (Claude Sonnet, parallel chunks), Spot Check (Claude Sonnet audit), and Validation.

Horizontal Formats

Standard one-row-per-Unit layouts where every field is directly visible in the source.

Vertical / Multi-Row Formats

Layouts where each Unit spans multiple rows (charge code + amount on separate sub-rows). The pipeline aggregates charges across sub-rows automatically.

Totals

Download Test Files

For each property, download the original rent roll and the standardized 40-field output. Each output includes a “Rent Roll” sheet (extracted data) and a “Validation Report” sheet (spot check results, warnings, and extraction metadata).

Warning Details

All 15 warnings are informational flags for human review — none indicate data integrity failures or extraction errors. Every warning falls into one of three categories:

Row Count Mismatch (10 warnings)

On vertical-format files, each Unit spans multiple source rows. The pipeline correctly aggregates sub-rows into one Unit record. The warning simply notes the difference between source row count and extracted Unit count.

Charge Sum Rounding (3 warnings)

When aggregating individual charges into total_charges, sub-penny amounts occasionally produce rounding differences of a few dollars. The individual charge values are correct.

Occupied with Zero Base Rent (2 warnings)

These are real data conditions in the source file — not extraction errors. Common for employee units, units in transition, or units with non-standard lease structures.

How the Pipeline Validates Quality

Five overlapping checks run on every file:

1. Structure Analysis — AI reads the spreadsheet layout guided by a data dictionary with 100+ known header aliases and charge code abbreviations across Yardi, RealPage, RentManager, and AppFolio systems.

2. Extraction — Parallel chunked AI extraction with overlap regions to prevent Unit boundary splits. Deduplication uses composite keys (property + unit ID).

3. Spot Check — Independent AI audit of 8 sampled Units compares raw source cells against extracted values. Samples are biased toward edge cases (first rows, last rows, vacant Units, random fill). Scoring starts at 100; each value mismatch deducts 15 points, each formatting difference deducts 5. Multi-row Units are fully supported — the auditor sees all source rows for each Unit, not just the first.

4. Validation — Programmatic checks: missing unit IDs, duplicates, charge sum integrity, occupied-Unit rent sanity.

5. Comparison View — Interactive UI shows source → canonical field mapping for every Unit, with transformation badges.

Vendor Format Coverage

The test suite covers 8 property management system export formats commonly encountered in multifamily acquisitions:

Canonical Schema

Every file is normalized to the same 40-field schema regardless of source format. See the full Canonical Schema reference.