Skip to main content
Extraction Accuracy

Multi-Pass Extraction Results from Real Leases

Five commercial leases. Nine CAM fields each. Verified against SEC EDGAR public filings — no synthetic data.

SEC EDGAR public filings · No synthetic data · Verified results

By Angel Campa · ·

How the Pipeline Works

Three specialized models in sequence — each with a different job.

1

Full Extraction

MiniMax M2.7

Reads the entire lease and extracts all 9 CAM fields in a single structured pass.

2

Adversarial Validation

GPT-5.4 Mini

Challenges Pass 1 output. Corrections require ≥70% confidence to override.

3

Escalation Tiebreaker

GLM-5

Fires on <20% of leases when Pass 1 and Pass 2 disagree on high-stakes fields.

Merge priority: Pass 2 corrections override Pass 1 only at ≥70% confidence. Pass 3 tiebreaker takes precedence over both when triggered.

9 fields

Per lease, every CAM reconciliation variable

~$0.50

Typical cost per lease processed

4–8×

Cost reduction vs. single-pass Claude Sonnet

Five Leases, Verified Results

Each lease sourced from SEC EDGAR public filings. Extraction results verified against lease text.

Stella Link Shopping CenterNeurogene

NNN Multi-Tenant
  • base_year
    2024
    Source: §4.1
  • base_year_amount
    Not stated
    Source: §4.1
  • gross_up_base_year
    95%
    Source: §4.1
  • pro_rata_share
    63.54%
    Source: §23.1
  • cap_type
    non-cumulative
    Source: §23.3
  • cap_rate
    5%
    Source: §23.3
  • admin_fee_percentage
    15%
    Source: §23.5
  • excluded_pools
    Capital, structural
    Source: §23.2
  • accounting_basis
    cash
    Source: §23.1
Multi-Pass Advantage

15% admin fee buried in §23.5

Single-pass models frequently miss admin fee clauses embedded deep in long retail leases. The adversarial pass confirmed the 15% rate with ≥70% confidence, preventing a common landlord under-collection error.

The 9 CAM Reconciliation Fields

Every field required to fully reconstruct a tenant's CAM exposure.

base_year

Calendar year used as the expense baseline for stops

base_year_amount

Dollar amount of the base year stop, if explicitly stated

gross_up_base_year

Occupancy percentage to normalize base year to full-occupancy costs

pro_rata_share

Tenant's percentage share of CAM expenses

cap_type

Whether expense cap is cumulative (compounding) or non-cumulative

cap_rate

Annual percentage ceiling on CAM expense increases

admin_fee_percentage

Management / administrative fee added on top of direct expenses

excluded_pools

Expense categories explicitly excluded from CAM (capital, structural, etc.)

accounting_basis

Whether expenses are tracked on cash or accrual basis

Pipeline Economics

Three-model architecture keeps costs low without sacrificing accuracy.

No disputes (typical)

Pass 1
~$0.30
Pass 2
~$0.15
Pass 3
Total
~$0.45

With escalation

Pass 1
~$0.30
Pass 2
~$0.15
Pass 3
~$0.10
Total
~$0.55

Worst case

Pass 1
~$0.50
Pass 2
~$0.25
Pass 3
~$0.20
Total
~$0.95

vs. single-pass Claude Sonnet: $2–4 per lease

4–8× savings
3-model fallback chains — no single point of failure
Exponential backoff retry on transient errors
Circuit breaker protection per provider

See Multi-Pass Extraction on Your Leases

Upload a lease PDF and get structured CAM fields in minutes.

Extract My Lease Fields
Protected customer data
Results in minutes
30-day free trial