The Six Principles

When MAS published the MindForge AI Risk Management trilogy in 2024, it did something unusual: it gave Singapore financial institutions a structured governance anchor rather than a principles-only framework. The 17 MindForge Considerations are not aspirational — they are the operational baseline MAS expects institutions to demonstrate.

This article extracts six governing principles from MindForge, maps the reinforcing guidance from ABS and IMDA, and shows where the international frameworks (SR 11-7, EU AI Act, NIST AI RMF, MAS FEAT, SS1/23, OECD) converge and diverge. The goal is a single reference taxonomy that a Line 2 practitioner can use to build a non-duplicative, cross-jurisdictionally coherent AI risk governance programme.

Why MindForge as the Anchor

There are good reasons to treat MindForge as the primary bible rather than a framework-neutral taxonomy:

Jurisdictional specificity. MindForge is written by, and for, MAS-regulated financial institutions. It reflects MAS’s supervisory experience and expectations in a way that generic frameworks like NIST AI RMF cannot. When MAS examiners ask about your AI governance, they are measuring against MindForge — not NIST.

Operational granularity. The 17 Considerations in MindForge are more operationally specific than most comparable frameworks. They tell you what to have in place, not just why it matters.

Three-document depth. The Executive Handbook sets governance expectations; the Operationalisation Handbook translates them into process; the Implementation Examples ground them in practice. Together they form a complete implementation guide.

Secondary reinforcement. The ABS GenAI Handbook and IMDA MGF for Agentic AI are explicitly aligned with MindForge. Using MindForge as the anchor means these secondary documents layer on top naturally.

The Six Principles

The 17 MindForge Considerations cluster into six governing principles:

P1 — Clear Human Accountability and Oversight

MindForge Considerations C1, C2

This is the most cross-framework-consistent principle in the taxonomy. Every major framework — MindForge, FEAT, SR 11-7, EU AI Act, NIST AI RMF — requires clear human ownership of AI systems and board-level oversight. The Three Lines of Defence model maps cleanly.

What MindForge adds is operational specificity: C1 requires named roles (not committee-level diffusion), and C2 requires that oversight mechanisms be designed before deployment, not retrofitted. The ABS Handbook reinforces this with concrete expectations for Board-level GenAI risk reporting.

International alignment: SR 11-7’s model ownership and independent review requirements translate directly. EU AI Act Article 17 (quality management system) and Article 9 (risk management) require comparable accountability structures for high-risk AI.

---

P2 — Proportionate, Risk-Based AI Governance

MindForge Considerations C3, C4, C5

This principle has no direct equivalent in the older framework-neutral taxonomy — it is genuinely MindForge-native. The requirement to develop an institution-specific AI risk taxonomy (C3), define an AI risk appetite (C4), and apply proportionate controls (C5) is a meaningful governance obligation that goes beyond what most international frameworks require at this level of specificity.

The contrast with the EU AI Act is instructive: the EU Act imposes a top-down legislative risk classification (unacceptable / high / limited / minimal). MindForge takes a bottom-up institutional approach — FIs must build their own risk taxonomies calibrated to their specific AI portfolios. Both converge on the principle that governance intensity must be commensurate with risk.

International alignment: NIST AI RMF’s MAP function (understanding the context and risk of AI systems) is the closest analogue. SS1/23 (MAS Technology Risk Management) provides complementary technology risk classification methodology. OECD Principle 1.2 (risk management) aligns at the high level.

---

P3 — Responsible Use Across the AI Lifecycle

MindForge Considerations C7, C10, C12, C13, C14, C15

This principle absorbs the substance of the prior GP-VALID-01 and GP-MONITOR-01 principles. MindForge’s lifecycle framing is more comprehensive: it begins with use-case definition (C7), covers independent model validation before deployment (C10), pre-deployment testing (C12), post-deployment monitoring (C13), change management (C14), and controlled retirement (C15).

SR 11-7 remains the gold standard for the validation dimension — its detailed requirements for independent review, challenger models, back-testing, and sensitivity analysis are more specific than MindForge’s C10. But SR 11-7 was written for statistical models; applying it to LLMs and agentic AI requires significant interpretive work. MindForge’s lifecycle framing is more fit-for-purpose for the full spectrum of modern AI.

International alignment: SR 11-7 (strong, validation focus). EU AI Act Article 9 (risk management system — covers monitoring). NIST AI RMF MANAGE and MEASURE functions. MAS FEAT Section 4 (monitoring expectations for AIDA systems).

---

P4 — Data, Model, and System Integrity & Soundness

MindForge Considerations C8, C9, C11, C17

This principle requires careful explanation because it absorbs three prior principles: GP-DATA-01 (data quality), GP-FAIR-01 (fairness), and the model-soundness dimension of GP-EXPLAIN-01 (model-level interpretability).

On fairness: MindForge has no standalone Fairness principle. This is a deliberate design choice — MindForge treats fairness as a data-and-model-integrity obligation (C8: ethical/compliant data; C9: data quality and bias controls), not a separate governance pillar. FEAT’s Fairness chapter (Section 2) maps strongly here. For practitioners: if your SR 11-7-based validation framework doesn’t include explicit bias testing methodology, you have a gap that both FEAT and MindForge expect you to fill — but the framing is “data quality and model soundness,” not “fairness compliance.”

On model interpretability: C11 addresses model-level explainability — the ability to understand how input features drive output predictions. This is a technical capability requirement, distinct from the obligation to explain decisions to affected individuals (which lives in P5).

International alignment: FEAT Section 2 (Fairness — strong). EU AI Act Article 10 (data governance for high-risk AI). NIST AI RMF MAP and MEASURE. MAS FEAT Section 4 (Accountability — data governance).

---

P5 — Transparency, Traceability, and Auditability

MindForge Considerations C6, C11 (decision-level), C15

This principle absorbs the prior GP-DOCUMENT-01 and the decision-transparency dimension of GP-EXPLAIN-01. The two dimensions are:

External transparency (C6): Disclosure to customers when AI is meaningfully involved in material decisions. Customers are entitled to know they are being assessed by an AI system, and to receive an explanation of the outcome commensurate with its impact on them.

Internal auditability (C15): Comprehensive documentation and audit trails — design rationale, validation findings, deployment approvals, change history — sufficient for supervisory review. The documentation standard: would a technically competent reviewer, with no prior knowledge of this system, be able to reconstruct and assess the governance process from the records alone?

International alignment: EU AI Act Article 13 (transparency for high-risk AI) and Article 12 (record-keeping). NIST AI RMF GOVERN function (documentation as risk practice). MAS FEAT Section 3.2 (transparency and explainability — now mapped here for the decision-transparency dimension).

---

P6 — Organisational Capability and Responsible AI Culture

MindForge Considerations C16, C17

This principle is genuinely new — no prior principle in the framework-neutral 7-principle structure covered organisational capability and culture explicitly. MindForge is unusual among governance frameworks in making culture a first-class governance obligation.

C16 requires documented, role-appropriate training for everyone with AI governance responsibilities — from Board members to model validators. C17 requires that senior leadership actively promotes a culture in which AI risks are surfaced without fear of reprisal. MAS assesses culture through thematic reviews and the quality of Board-level AI risk reporting, not just documentation.

The ABS Handbook’s training requirements reinforce this directly. Notably, SR 11-7 does not address culture — it is a technical model risk standard, not an organisational governance framework. If your AI governance programme is SR 11-7-anchored, this is a dimension you need to build from scratch using MindForge and IMDA guidance.

International alignment: ABS GenAI Handbook (strong — training requirements). IMDA MGF for Agentic AI (moderate — responsible deployment principles). OECD Principle 4 (education and capacity building). No direct equivalent in SR 11-7.

---

The Interactive Knowledge Graph

The graph below maps the six principles (hexagons) against the 14 source documents (circles). Line weight and colour reflect alignment strength. Click any node to see its connections.

PrincipleSG DocumentIntl DocumentClick · Drag

Cross-Framework Alignment at a Glance

The matrix below shows which frameworks provide strong, moderate, or no coverage for each principle. Singapore-jurisdiction documents appear on the left; international frameworks on the right.

Principle	MAS FEAT	MAS AIRM	MAS Info Paper	MAS MindForge Exec	MAS MindForge Ops	MAS MindForge Impl	ABS GenAI	IMDA MGF	EU AI Act	Fed/OCC 11-7	NIST AI RMF	PRA SS1/23	ISO 42001	OECD AI
Accountability & Oversight C1, C2	●	●	—	●	●	—	●	●	●	●	●	●	●	●
Risk-Based Governance C3, C4, C5	◐	●	—	●	●	—	●	●	●	◐	●	◐	●	◐
Lifecycle Responsibility C7, C10, C12, C13, C14, C15	◐	●	—	●	●	—	●	●	●	●	●	●	◐	◐
Integrity & Soundness C8, C9, C11, C17	●	●	—	●	●	—	●	●	●	●	●	●	◐	●
Transparency & Auditability C6, C11, C15	●	●	—	●	●	—	●	●	●	●	●	●	◐	●
Capability & Culture C16, C17	—	◐	—	●	●	—	●	●	—	—	◐	—	●	◐

● Strong alignment◐ Moderate alignment— Not addressedSG = Singapore frameworksIntl = International frameworks

Three things to notice:

P2 (Risk-Based Governance) has the sparsest international coverage. Most frameworks assume you have a risk taxonomy; only MindForge and the EU AI Act give substantive guidance on building one. This is the principle where following MindForge most closely produces governance value that international frameworks don’t provide.

P6 (Culture) is almost exclusively MindForge, ABS, and IMDA territory. SR 11-7 is silent. EU AI Act touches it only obliquely. If you work at a bank that runs its AI risk programme off SR 11-7, culture is your biggest framework gap.

P1 (Accountability) shows the strongest cross-framework alignment — every framework agrees that clear human ownership and board oversight are non-negotiable. This is the safest place to start if you’re building a governance framework from scratch.

What This Means in Practice

For Line 2 practitioners building or reviewing AI governance frameworks:

Use the six principles as your control taxonomy. Map every policy, procedure, and control to a principle — not to a specific regulatory document. When regulators ask “how do you address auditability?”, your answer draws from MindForge C15, FEAT Section 3.2, EU AI Act Article 12, and SR 11-7 model documentation requirements simultaneously. This is the architecture of a mature, non-duplicative governance programme.

On the fairness subsumption: If you have existing Fairness policies mapped to GP-FAIR-01, reclassify them under P4. Your FEAT fairness obligations still apply — nothing is lost. What changes is the framing: you are managing fairness as a data-quality and model-integrity problem, which is both more technically grounded and more aligned with how MindForge and MAS supervisors think about it.

On the lifecycle principle (P3): If your model risk management framework is SR 11-7-based, validate it covers all of MindForge C7–C15. SR 11-7 is strong on independent validation; it is weaker on use-case definition (C7), pre-deployment testing (C12), and model retirement (C15). These are your SR 11-7 gaps relative to MindForge expectations.

---

This is the first article in the Principle Mapping series. Next: a deep dive into MindForge Considerations C12–C15 and the lifecycle governance of agentic AI — where the IMDA MGF guidance becomes the critical secondary source.

Explore the full interactive knowledge graph at principle.sg/graph.