Language models hallucinate. In creative applications, this is a feature — it's what makes AI-generated text feel human. In mortgage income calculation, it's a liability. A deterministic income engine eliminates this risk by using published Fannie Mae 1084 guidelines as executable code — same inputs, same output, every time. At Confer, we deploy AI where it excels (document classification and data extraction) and deterministic math where correctness is non-negotiable (income calculation, DTI, APR).
The Hallucination Problem in Mortgage Math
When a language model calculates self-employed income from a Schedule C, it's performing a task it was never designed for. LLMs are pattern-matching engines trained on text. They can describe how to calculate Schedule C income with impressive accuracy. But when asked to actually perform the calculation — add back depreciation on Line 13, add back 50% of meals on Line 24b, apply the amortization adjustment from Form 4562 — they produce results that vary between runs.
This isn't a bug. It's the fundamental architecture of probabilistic text generation. The model samples from a distribution of likely next tokens. Most of the time, the math works out. Some of the time, it doesn't. And in mortgage underwriting, "some of the time" is unacceptable.
What the Data Shows
According to ACES Quality Management's Q2 2025 report, the overall critical defect rate in mortgage origination reached 1.51% — up from a record low of 1.16% in Q4 2024. Income and employment defects led all categories at 25% of total critical defects, the leading category for the fifth consecutive quarter.
What "Deterministic" Actually Means
A deterministic system produces the same output for the same input, every time. There is no randomness, no sampling, no temperature parameter, no probability distribution. The function is a mathematical transformation — reproducible, testable, and verifiable.
Run this calculation today, tomorrow, or next year. The answer is $8,816.67. Run it ten thousand times. Still $8,816.67. There is no variance because there is no probabilistic component.
The Hybrid Architecture: AI + Deterministic Math
The question isn't "AI or no AI." It's "AI where, and deterministic math where."
Where AI Excels
- •Document classification — Pattern matching handles 70%+ of cases, LLM handles ambiguous documents
- •Data extraction — Pulling numbers from W-2 Box 1 or Schedule C Line 31
Where Deterministic Code Is Required
- •Income calculation — Fannie Mae 1084 guidelines implementation
- •DTI computation — Total obligations ÷ qualifying income
- •APR derivation — 12 CFR 1026.4 compliance
The Seven Income Calculators
Confer implements seven income type calculators, each following the specific Fannie Mae guideline for that income source:
W-2 Salary (Fannie Mae B3-3.1)
YTD earnings plus 2-year W-2 history. Cross-checks pay stub amounts against W-2 totals.
Self-Employment: Schedule C (Fannie Mae 1084)
Net profit plus add-backs for depreciation, amortization, business use of home, and 50% of meals.
Rental Income: Schedule E (Fannie Mae B3-3.1-05)
Gross rents adjusted by 75% vacancy factor, minus operating expenses, plus depreciation add-back.
K-1 Partnership/S-Corp (Fannie Mae 1084)
Pass-through income with entity-type-specific adjustments. Partnership vs S-Corp rules differ.
Investment Income
Dividend and interest trending across 2-year history with declining pattern detection.
Retirement Income
Social Security, pension, annuity, IRA distributions. Validates 3-year continuity requirement.
Other Income
Alimony, child support, VA benefits, disability. Verifies documentation and continuity.
The Audit Trail: Why It Matters
Every calculation produces a worksheet matching the Fannie Mae 1084 format. The worksheet shows:
- Which source documents were used (specific W-2s, tax returns, pay stubs)
- The calculation method applied (guideline reference)
- Each add-back with its amount, source line, and justification
- 2-year trending analysis with year-over-year comparison
- Any flags or review items with specific reasons
- The final qualifying monthly income
Repurchase demands don't come with a "your AI was right most of the time" defense. They come with a question: "Show me how you calculated this borrower's income." A deterministic engine with a full audit trail answers that question completely.
Implementation: Where the Boundary Lives
The architecture makes the boundary between AI and deterministic processing explicit:
The double line in the middle is the boundary. Above it, AI processes documents. Below it, deterministic code does math. The underwriting analysis at the bottom uses AI assistance (risk scoring, condition recommendations) — but the income numbers feeding into that analysis were calculated deterministically.