AI-MRG Framework
An open-source, full-lifecycle AI and model risk governance toolkit for U.S. regulated industries — starting with financial services.
Status: v0.3 (Phase 3 complete). New in v0.3: a web dashboard, MLflow integration, Docker Compose, the generative/agentic AI track (evaluation, red-team and agent-permission checks mapped to NIST AI 600-1), Colorado SB 26-189 controls and draft notices, and verifiable examination packages. From v0.2: Inventory, classification, risk tiering, control planning, automated checks (data, performance, stability, outcomes, fair lending, adverse-action reasons), a gated lifecycle workflow with independent approval, a tamper-evident evidence trail, findings with aging, reports in Markdown and Word, a REST API with role-based access, PostgreSQL support and signed webhooks. Next: practitioner review, pilots and the 1.0 hardening (Phase 4). See
docs/ROADMAP.md. Pack content has not yet been reviewed by practitioners.
Why this exists
U.S. guidance tells institutions what good AI and model governance looks like. Turning that into day-to-day practice is the hard part. That means inventories, risk tiers, validation evidence, fairness testing, adverse-action reasons, monitoring thresholds and audit-ready documents. Enterprise platforms do this well, but they are licensed products. Community banks, credit unions, fintech start-ups and academic researchers often cannot justify the cost.
The 2026 regulatory reset makes the gap sharper:
- SR 26-2 / OCC Bulletin 2026-13 (April 17, 2026) replaced SR 11-7. It is principles-based and risk-based, sets no enforceable standards and is aimed mainly at banks over $30B in assets. It explicitly excludes generative and agentic AI. Institutions are told to cover those with their own governance practices.
- ECOA / Regulation B still require specific, accurate adverse-action reasons, whatever the model's complexity (CFPB Circulars 2022-03 and 2023-03).
- Colorado SB 26-189 brings automated-decision-making obligations for lending and other consequential decisions starting January 1, 2027.
- NIST AI RMF 1.0 and its Generative AI Profile (NIST AI 600-1) remain the main voluntary reference. A revision is in progress.
AI-MRG turns these sources into versioned control catalogs, automated checks, evidence trails and document templates that any institution can run locally, with no license fee, no vendor account and no telemetry.
What it will do
| Capability | Description |
|---|---|
| Inventory & classification | Register every model and AI system. Classify it as model under SR 26-2, or as GenAI or agentic AI, and route it to the right governance track. |
| Risk tiering | Configurable scoring on inherent risk, exposure, purpose, materiality and aggregate risk, using SR 26-2's risk factors. |
| Control catalog | Lifecycle controls stored as data (YAML), each with its regulatory references, required evidence and applicability rules. |
| Checks library | Python checks for data quality, performance, stability, outcomes analysis, fair lending and adverse-action reason quality. |
| Evidence & audit trail | Append-only, hash-chained evidence records linked to controls, findings and approvals. |
| Document generation | Development documents, validation reports, monitoring reports, adverse-action reason maps and inventory reports. |
| Dashboard | A web UI for inventory, review workflows, findings and monitoring. |
What it is not
- Not legal advice, and not a certification. Using AI-MRG does not by itself make an institution compliant.
- Not a replacement for independent validation or human judgment. It structures and evidences that work.
- Not a hosted service. You run it yourself.
Quick start
git clone https://github.com/abiroland/ai-mrg-framework
cd ai-mrg-framework
pip install -e ".[dev]"
# A complete run on a simulated credit model: registry, evidence, findings, reports
python examples/credit-scorecard/run_example.py demo
Or run PostgreSQL, the API and the dashboard together:
POSTGRES_PASSWORD='choose-one' docker compose up -d --build
docker compose exec api cat /data/bootstrap-admin-token # sign in at http://127.0.0.1:8501
Then follow Getting started for the step-by-step workflow (aimrg init → register → classify → tier → plan → check run → report render → audit verify). See Checks and Writing packs for details.
Documentation
- Architecture: system design, modules, data model, interfaces
- Regulatory map: how controls trace to SR 26-2, ECOA/Reg B, NIST AI RMF, Colorado and others
- Roadmap: phased delivery plan and success metrics
- Architecture decision records
Scope by release
- Financial services (U.S. banking and consumer lending): first, v0.1–v1.0
- Healthcare: planned as a second sector pack on the same core, after v1.0
Contributing
See CONTRIBUTING.md. Regulatory content changes must cite a primary source.
License
Apache License 2.0. See ADR-0001.
Metadata
Release files for ai-mrg 0.3.0
For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.
Source distribution (sdist)
| File | Size | Uploaded | Provenance | Actions |
|---|---|---|---|---|
| ai_mrg-0.3.0.tar.gz | 653.3 kB | With provenance | Details for ai_mrg-0.3.0.tar.gz |
Built distribution (wheel)
| File | Interpreter | ABI | Platform | Actions Reset |
|---|---|---|---|---|
| ai_mrg-0.3.0-py3-none-any.whl | Python 3 | none | any | Details for ai_mrg-0.3.0-py3-none-any.whl |
Total release size: 787.3 kB
Release files / ai_mrg-0.3.0.tar.gz
| Download URL | ai_mrg-0.3.0.tar.gz |
|---|---|
| Size | 653.3 kB |
| Tags | Source |
|
SHA-256 checksum How to use checksums |
df2bd37de82f66b22898acf0c4be415b06a0e0cd1d32dc86d387a53d675d6436
|
|
BLAKE2b-256 checksum How to use checksums |
c976647aa76794553ccd3053ae9420daa56458e70bf6fb8242c5ab728f9c73eb
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
Yes |
| Uploaded via |
twine/7.0.0 CPython/3.13.16
|
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
PyPI Publish Attestation
PyPI verified that this artifact, at this checksum, originated from the publisher listed below.
Signed by GitHub Actions, verified by PyPI on Nov 10, 2026.
Transparency logRelease files / ai_mrg-0.3.0-py3-none-any.whl
| Download URL | ai_mrg-0.3.0-py3-none-any.whl |
|---|---|
| Size | 134.0 kB |
| Tags | Python 3 |
|
SHA-256 checksum How to use checksums |
695afbaf354ffb6b04d018c2765a39708bc228313a5e957b7478f74b4cd909df
|
|
BLAKE2b-256 checksum How to use checksums |
423d8f126d58742a4fb4c0b940abdbf6456f5a5564fd3d18e730339da6551b00
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
Yes |
| Uploaded via |
twine/7.0.0 CPython/3.13.16
|
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
PyPI Publish Attestation
PyPI verified that this artifact, at this checksum, originated from the publisher listed below.
Signed by GitHub Actions, verified by PyPI on Nov 10, 2026.
Transparency logSecurity for ai-mrg 0.3.0
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
Full provenance
Both files in this release have been uploaded with provenance metadata.
Attestation for ai_mrg-0.3.0.tar.gz
PyPI Publish Attestation
PyPI verified that this artifact, at this checksum, originated from the publisher listed below.
Signed by GitHub Actions, verified by PyPI on Nov 10, 2026.
Transparency logA note on package security
PyPI is a repository for distributing packages, not a service for auditing them. While third-party security researchers and our staff do their best to detect and remove malicious packages, PyPI cannot guarantee the quality, security or suitability of any given package.
To learn more about protecting yourself from malicious actors, visit our package security documentation.