Gemini 2.5 Pro Matrix Analysis: Montage Matrix v 1.0 (01-01-2026)

Excellent. As a senior mortgage compliance and underwriting guidelines analyst, I have completed a thorough review of the provided FundLoans "Montage (Full Doc)" eligibility matrix and its corresponding if-then rules.

Here is my comprehensive analysis, structured according to your five key focus areas.

Executive Summary

The auto-generated if-then rules are a good first step but contain critical flaws that prevent them from being reliably used for underwriting or a customer-facing chatbot. While the direct data transcription is mostly accurate, the rules fail to handle ranges and boundary conditions correctly, contain significant gaps in coverage, and have organizational inconsistencies. The most severe issue is the misinterpretation of FICO score tiers as discrete points rather than ranges, rendering the rules unusable for any FICO score not explicitly listed.

1. Rule Accuracy

The transcription of specific LTV percentages, loan amounts, and other values from the markdown tables into the if-then rules is largely accurate. However, there are a few notable errors and omissions:

• Correct Transcription: For a given combination present in the markdown table (e.g., Owner Occupied, >$1.5 MM to $2 MM, 700 FICO), the corresponding LTVs in the if-then rule (80% Purch/RT, 75% CO) correctly match the source. The translation of "N/A" to "Not Eligible" is also accurate and helpful.

• Minor Inaccuracy (Omission): The "Assets Only / Asset Allowance" section in the markdown specifies "Max LTV 80%" and "Max 50% DTI". The corresponding if-then rule omits this information, stating only, "See Guidelines for Formulas*". This is a failure to capture all explicit rules from the source.

• Organizational Inaccuracy: Under the "Second Home" section, a rule for Prepayment Penalties states: Minimum of 1 year prepay required on investment properties. This rule belongs in the Investment property section, not the Second Home section. Placing it here is misleading and will cause retrieval errors.

2. Edge Cases & Ambiguity

This is the area with the most significant flaws. The rules are not robust enough to handle boundary conditions, which would lead to frequent "no-match" scenarios.

• Critical Flaw (FICO Tiers): The rules define FICO scores as exact values (e.g., IF...FICO is 740). Mortgage guidelines almost always work in tiers (e.g., 740+, 720-739). A borrower with a 739 FICO would not match the "740" rule or the "720" rule, making them seem ineligible.

  • Recommendation: The rules must be rewritten to use ranges or "greater than or equal to" logic. For example: IF...FICO is >= 740... or IF...FICO is between 720 and 739....

• Loan Amount Boundaries: The handling of loan amount boundaries is better but could be clearer. The use of up to $1 MM and >$1 MM to $1.5 MM is logically sound, implying that a loan of exactly $1,000,000 falls into the first category. This is standard industry practice but could be made more explicit (e.g., loan amount <= $1 MM).

• Ambiguous Note: The note "Not allowed L/A >$3MM" under "Other LTV/CLTV Limits" for all occupancy types is highly ambiguous. It directly contradicts the "Super Jumbo" sections, which explicitly provide rules for loans up to $6MM. This note likely applies only to the property types in that specific table (e.g., Condos, 2-4 units), but its placement and wording create a major policy conflict.

3. Missing Scenarios (Gaps in Logic)

Several combinations of loan parameters are not covered by the rules, creating significant gaps.

• Major Gap (Owner Occupied Super Jumbo): There are no rules for Owner Occupied loans between $3.5 MM and $4 MM. The matrix jumps from ">$3 MM to $3.5 MM" directly to ">$4 MM to $5 MM". This is a critical gap.

• Missing FICO Tiers: For Second Home and Investment properties in the "up to $1MM" loan tier, the highest FICO listed is 700 and 740, respectively. While it's implied that a 760 FICO would receive the best terms available (the 700 or 740 tier), an explicit rule is missing, which could cause lookup failures.

• Missing Reserve Rules: The Owner Occupied and Investment sections have clear reserve requirement rules. The Second Home section has no rules for reserves, a major omission. A chatbot or underwriter would not be able to determine the reserve requirement for a Second Home.

• Missing Eligible Borrower Rules: The Investment section lists eligible borrower types. The Owner Occupied and Second Home sections do not. It is unclear if the eligibility is different or if it was simply omitted. This is a crucial compliance and eligibility point.

4. Cross-Product Consistency

Comparing the rules across Owner Occupied, Second Home, and Investment properties reveals logical consistencies as well as some glaring inconsistencies.

• Logical Consistency: The overall LTV structure is consistent with industry risk-based pricing. Owner Occupied properties generally receive the highest LTVs, followed by Second Homes, and then Investment properties, which is expected.

• Inconsistency (Super Jumbo Gaps): The missing loan amount range of $3.5MM to $4MM is unique to the Owner Occupied section. The Second Home and Investment Super Jumbo sections correctly cover a ">$3 MM to $4 MM" range. This inconsistency is a likely error in the Owner Occupied matrix.

• Inconsistency (Feature Availability): Key guidelines like Reserves, Eligible Borrowers, and Prepayment Penalties are not presented consistently across all three occupancy types, as detailed in the "Missing Scenarios" section above. This suggests the source documentation itself may be incomplete or poorly organized.

5. Chatbot Readability

Rating: 2.5 / 5 (Poor to Fair)

While the self-contained, IF-THEN format is a good foundation for RAG (Retrieval-Augmented Generation), the current implementation has severe limitations that would result in a frustrating and often incorrect user experience.

• Strengths:

  • The atomic, self-contained structure is ideal for semantic search and retrieval.
  • The use of consistent fields (program, occupancy, FICO, loan amount) provides a clear structure for queries.

• Weaknesses:

  • The FICO Tier Flaw: This is the primary reason for the low score. A chatbot would fail to answer the vast majority of user queries (e.g., "What's the max LTV for a 725 FICO?"). It could only answer for the exact FICO scores of 660, 680, 700, etc.
  • Unfriendly THEN Clauses: Many rules resolve to non-natural language outputs. For example, THEN Housing (12 mos. Minimum) = 0x30x12. A chatbot would need an additional translation layer to explain this to a user as "You must have zero 30-day late payments on your mortgage in the last 12 months."
  • Gaps & Ambiguity: The missing scenarios and ambiguous notes would force the chatbot to frequently respond with "I don't have enough information" or, worse, provide a confident but incorrect answer based on conflicting rules (e.g., the ">$3MM" loan amount conflict).
  • Organizational Errors: The misplaced Prepayment Penalty rule would cause the chatbot to give incorrect information about PPPs for both Second Homes and Investment properties.

Conclusion & Final Recommendations

Before these rules can be used in a production environment, they require a significant overhaul:

1. Rewrite FICO conditions to use ranges or "greater than/equal to" logic.
2. Fill all identified gaps, especially the Owner Occupied Super Jumbo loan amount range and the missing Second Home reserve rules.
3. Clarify ambiguous rules, particularly the "Not allowed L/A >$3MM" note, by adding the necessary context.
4. Reorganize misplaced rules (e.g., move the Investment PPP rule to the correct section).
5. Standardize the THEN clauses to be more descriptive and human-readable to reduce the burden on the chatbot's interpretation layer.
6. Add rules for missing information across all products, such as "Eligible Borrowers".