The AI Challenge: The "Black Box" in Banking

Traditional credit scoring is hitting a wall. As financial institutions move toward automated decision-making, they face a "Trust Gap." Standard AI models are probabilistic—they offer a "best guess" based on historical data, but they often struggle to explain why a loan was denied.

In a regulatory climate governed by Quebec’s Law 25 and OSFI Guideline E-23, "I don't know" is no longer an acceptable answer from an algorithm. The QGI Loan Assessment Engine was built to solve this. It shifts the paradigm from guessing creditworthiness to proving alignment through a formal constraint system.

The Solution: Deterministic Assessment

The Loan Assessment Engine doesn't just look at a number; it looks at the business. By using a Portfolio-Based Model, the engine evaluates a business owner’s real-world reasoning, situational judgment, and operational resilience.

This isn't just "faster" lending; it is Governed Lending. Every step of the evaluation is filtered through the QGI 4-Tier Architecture, ensuring that the AI operates within strict mathematical boundaries that it physically cannot break.

The Architecture: How The System Works

The QGI architecture separates policy from execution, ensuring that safety is baked into the kernel of the system.

Tier 1: Policy Configuration (The Risk Compiler)
Before the assessment begins, Tier 1 establishes the "Safety Profile" based on the loan's risk class (e.g., Micro-loan vs. Commercial Expansion).

• Role: Sets the strictness constants (τ) for downstream invariants.
• Banking Application: Ensures that for high-impact loans, the Non-Harm and Opacity-Limit thresholds are automatically tightened to meet institutional risk appetite.

Tier 2: The 5 System Invariants (The Kernel)
The core of the engine. These five mathematical gates monitor every interaction between the AI and the applicant data.

1. Non-Harm (INH​): Physically blocks the use of prohibited bias vectors (e.g., gender, race, or proxy-data) in the decision logic.
2. Autonomy (IA​): Enforces strict data-sharing boundaries; the applicant "unlocks" specific portfolio artifacts (bank feeds, tax returns) for a limited evaluation window.
3. Opacity-Limit (IOL​): Requires the AI to maintain a "Reasoning Trace." If a recommendation cannot be logically explained in human-legible terms, the engine blocks the output.
4. Mutual Benefit (IMB​): Ensures the depth of the assessment is proportional to the loan value, preventing "information extraction" from the small business.
5. Evolvability (IE​): Continuously monitors for "Sector Drift," alerting administrators if the system begins favoring specific industries without a policy mandate.

Tier 3: Jurisdictional Profile (The Compliance Wrapper)
This layer maps the engine’s behavior to local legal requirements.

• Quebec/Canada Focus: Includes specific modules for Quebec’s Law 25 (Right to Explanation) and OSFI Guideline E-23.
• Adaptability: Tier 3 allows the engine to be "hot-swapped" for different regions (e.g., EU or US) without altering the core Tier 2 safety kernel.

Tier 4: Operational Preflight (The Action Gate)
The final execution layer where the applicant interacts with the system.

• Methodology: Applicants provide "Portfolio Artifacts" and respond to Scenario-Based Simulations (e.g., "How would your cash flow handle a 20% increase in interest rates?").

System Parameters & Specifications

The Output: The Business Alignment Map

We are retiring the "Credit Score." In its place, the engine generates a Business Alignment Map. Instead of a single, binary "Pass/Fail," lenders receive:

• Reasoning Patterns: How the applicant handles financial trade-offs.
• Capability Profile: A deep dive into operational strengths and growth areas.
• Transparency Disclosure: A plain-language explanation of the decision factors, ready for the member or regulator.

This shift preserves Human Decision Authority. The AI provides the verified evidence; the banker makes the informed choice.