Smoke test — zero LLMs

// RATE THIS DRAFT:

Title

// IDEA

// RESEARCH (Phase 2 will auto-generate)

# Research Document: HEI Eligibility Engines — No LLMs in the Decision Graph

---

## Context

Home equity investment origination is in an operational scaling moment — Point, Unlock, Hometap, and a growing list of regional operators are competing for the same constrained originator capacity while deal volume pressure mounts. Eligibility determination is the first hard gate: get it wrong early and you've burned underwriter time on files that never had a chance. The AI-in-fintech narrative right now skews heavily toward LLMs, but HEI eligibility is a deterministic, rules-dense problem — combined LTV thresholds, property type exclusions, FICO floors, state-specific regulatory carve-outs — where probabilistic token prediction is the wrong tool and operators are starting to figure that out.

---

## Supporting Facts

- Hometap, Point, and Unlock collectively operate in 15–25 states each, meaning any shared eligibility infrastructure must handle materially different ruleset stacks by jurisdiction — state-level overlays are a first-class engineering problem, not an afterthought. [unverified — confirm exact state counts per operator from their public product pages]
- LLM hallucination rates on structured rule-following tasks remain a documented failure mode; a 2023 Stanford HAI report and multiple legal/compliance evals found error rates high enough to be disqualifying for binary pass/fail decisioning workflows. [unverified — confirm specific Stanford HAI citation; the general finding is well-documented across multiple 2023–2024 evals]
- Most HEI operators still quote 30–60 day origination cycle times; eligibility pre-screening is one of the few stages where deterministic automation can compress the funnel without touching underwriter judgment. [unverified — confirm current cycle time benchmarks; figure is directionally consistent with public operator commentary]
- Rule engines (Drools, Blaze Advisor, or custom DSL implementations) have been the backbone of mortgage eligibility decisioning for 20+ years — the pattern is proven; the HEI-specific adaptation is new. [generally accepted; no single citation needed]
- CFPB has signaled ongoing scrutiny of algorithmic decisioning in credit-adjacent products; a deterministic, auditable rule graph is substantially easier to defend in an exam than a model whose weights you can't explain. [unverified — confirm current CFPB guidance language on algorithmic credit decisioning; directionally accurate based on 2023–2024 supervisory priorities]

---

## Related Angles (3)

1. **The auditability argument**: Regulators and counterparties both want to see the decision logic — a deterministic eligibility graph produces a decision trace; an LLM produces a vibe.
2. **Build vs. buy for origination tooling**: Most HEI shops are still duct-taping their eligibility checks inside their LOS or a spreadsheet — the case for a purpose-built eligibility layer versus extending a general-purpose LOS.
3. **Where LLMs actually belong in HEI ops**: Document extraction, borrower-facing Q&A, condition clearing — the contrast between appropriate and inappropriate LLM use cases in the same origination stack.

---

## Hook Options

- **Contrarian:** Everyone is shipping LLMs into their decisioning stack. We spent three months deliberately keeping them out — here's what that choice looks like in production.
- **Stat-led:** HEI operators run eligibility across 15–25 state rulesets per product. That's not a language problem. That's a logic problem.
- **Story-led:** First call with a new HEI client, they asked if we were using AI for eligibility. We said yes. Then we explained what kind — and it wasn't what they expected.
- **Contrarian (blunt):** LLMs don't belong in your eligibility graph. Not because they're new. Because they're probabilistic and your ruleset isn't.
- **Process-led:** Three months. Four operators. One rule: if the eligibility decision can't be traced to a specific line of logic, it doesn't ship.

---

## Closing CTA Options

1. If you're an HEI operator rebuilding your eligibility layer and want to talk about what a deterministic decision graph actually looks like in production — advisorylabs.xyz/contact.
2. Curious what's actually inside the decision graph of HEI shops you compete with — are they running rules engines, model scores, or something held together with spreadsheet logic and hope? Drop your take below.

// DRAFT (Phase 2 will auto-generate)

// FINAL — what you'll actually post

966 chars

// CAROUSEL SLIDES · optional · LinkedIn document post

7 slides

Download compiles + saves. LinkedIn → start a post → Document → upload the PDF.