CONTROLLED REFERENCE DEMONSTRATOR

North Star Mortgage

A bounded-AI governance case study for community-bank mortgage review

FINAL CONTROLLED STATE

AWAITING EVIDENCE AND QUALIFIED HUMAN REVIEW; NO CREDIT DECISION

Executive summary

The North Star Mortgage demonstrator shows how a synthetic community-bank mortgage case can be evaluated without allowing software - or an AI assistant - to make a credit decision. It separates deterministic calculation, regulatory-source retrieval, organizational authority and bounded AI operation while preserving a viable manual path.

Live demonstrator: roi-ea-integrated-demo-v01.vercel.app

1. The controlled case

The case is entirely synthetic. It represents a first-time home-purchase application whose financial indicators are close enough to fictional policy boundaries to require evidence review rather than an automated conclusion. Protected-class data and applicant age are excluded from the execution projection. The fictional policy is not attributed to any real bank, investor, regulator or mortgage product.

Measure

Calculated result

Controlled interpretation

Credit score

725

Within the fictional standard

Combined loan-to-value ratio

94.19%

Within the fictional standard

Total debt-to-income ratio

44.68%

Within a fictional exception-review band

Post-closing reserves

1.54 months

Within a fictional exception-review band

Unresolved evidence. Overtime-income continuity evidence has not been accepted, and a required current asset statement is missing. The demonstrator therefore reports insufficient evidence, identifies a manual exception-review candidate and abstains from a decision.

2. Four distinct governance layers

ROI-EA calculates permitted financial measures, traces inputs and compares them with fictional policy. It does not approve, deny, determine eligibility, price, counteroffer, issue a notice or establish realized ROI.

ERIR attempts read-only retrieval of controlled regulatory-source identifiers. Record retrieval is evidence of retrieval - not applicability, legal sufficiency, compliance or control effectiveness.

FACEM preserves the differences among access, evidence, recommendation, authority, acceptance, accountability and organizational commitment. No technical connection or imported workbook creates bank authority or commitment.

BACRM constrains the Mortgage Evidence Readiness and Consistency Assistant to deterministic calculations, evidence-gap detection, trace preparation and abstention. Consequential credit actions are prohibited.

3. Controlled spreadsheet ingestion

Users may download and complete the supplied MTG-IMPORT-V0.2 workbook. Import is fail-closed: the browser accepts only the four approved sheets, sentinel, headers and controlled fields. It rejects formulas, macros, external components, extra sheets, unknown columns, protected fields, unsafe URLs and invalid identifiers before changing the active case. Accepted data remain in browser-session memory and cannot create credit or action authority.

4. What the demonstrator establishes - and what it does not

The implementation demonstrates inspectable separation of calculation, source retrieval, organizational authority and bounded AI operation. It preserves a viable manual path and a pending human-disposition record.

It does not establish approval, denial, eligibility, pricing, fair-lending performance, legal compliance, model validity, production safety, operational effectiveness or realized return on investment. Federation value (C2 - C1) and bounded-AI value (C3 - C2) remain separately identified but unquantified because no member-specific operating evidence or measured AI-performance evidence has been supplied.

5. Verification status

Automated suite: 57 passed, 0 failed

Production spreadsheet import: manually confirmed

Production integrated trace: independently confirmed

Production implementation commit: 9c0d872

6. Public source context

These authoritative sources are candidates for qualified review. Their inclusion does not determine applicability:

Regulation B

HMDA / Regulation C

Fair Housing Act lending provision

FDIC community-bank third-party risk guide

NIST AI Risk Management Framework

© 2026 David C. Jones. All rights reserved.

AI at Human Scale