01 · FEAT · FOUR PRINCIPLES
Fairness · Ethics · Accountability · Transparency.
FS firms must demonstrate Fairness, Ethics, Accountability, and Transparency in AI/ML systems. Lifecycle controls span design, development, deployment, monitoring, and decommissioning. Per-decision evidence is the operative artefact a MAS examiner pulls.
MAS · FEAT principles + AIRM (in finalisation)
FEAT has been operative since 2018. AIRM formalises the principles into a lifecycle-aware risk regime. The Veritas Initiative is the industry-consortium toolkit that operationalises FEAT in code; the GitHub project ships open-source diagnosis and assessment tools in Python and Java. For the principle-by-principle reading, see MAS FEAT and AIRM, read against the AI agent.
F · Fair
Adverse-impact metrics per decision class.
WARRANT · trace.fairness_metrics (when supplied). Explainability rationale captured per action.
E · Ethics
Purpose alignment evidenced per action.
WARRANT · authorization_envelope.within_purpose per-action assessment.
A · Acct.
Clear owner of every AI-driven decision.
WARRANT · trace.signed_off_by + the record names the issuing tenant. Senior-officer binding when supplied.
T · Trans.
Explainability surfaced to data subject and supervisor.
WARRANT · trace.actions[*].rationale + per-action authorization_envelope.justification (immutable in the evidence record).
02 · LIFECYCLE SCOPE
Design through decommissioning.
AIRM extends FEAT across the full lifecycle. Design, development, deployment, monitoring, and decommissioning each carry evidence obligations. Per-trace authorization is assessed live; cross-trace lifecycle roll-up ships v0.5, 2026 Q3.
5stages
LIFECYCLE PHASES
Design, development, deployment, monitoring, decommissioning. Each phase carries FEAT obligations.
2018
FEAT IN FORCE
FEAT principles operative since 2018. AIRM is the lifecycle-aware formalisation under finalisation in 2026.
"Per-decision evidence is the operative artefact a MAS examiner pulls. Not the policy. The trail."MAS examiner reading · industry consortium · 2026-04-30
03 · VERITAS · INDUSTRY REFS
Open-source toolkit.
Veritas Toolkit v2.0 ships fairness diagnosis and assessment tools for Singapore FS. The GitHub repo is the canonical operational reference for FEAT in code. Allen and Gledhill, Swiss Re, and the Veritas open-source repo are the working primary sources while AIRM finalises.
04 · WHY THIS REGULATOR NOW
When did Singapore's FEAT AI principles take effect?
Singapore established FEAT in November 2018 as the first jurisdiction-level AI principles set issued by a financial-services regulator. The principles · Fairness, Ethics, Accountability, Transparency · have been operative across MAS-licensed institutions for seven years. AIRM (Artificial Intelligence Risk Management) is the lifecycle-aware formalisation now in finalisation under MAS's broader Information Paper on AI Model Risk Management. The regulator's working position is that FEAT is the principles set; AIRM is the operating manual.
Recent supervisory signal carries the framework's posture. The MAS Information Paper on AI Model Risk Management (December 2024 consultation, finalisation expected 2026 H2) sets the operational expectations for SG-licensed FIs deploying AI/ML in customer-facing decisioning. The Veritas Initiative (industry consortium led by MAS) released Veritas Toolkit v2.0 in November 2024, with HSBC, UOB, OCBC, and Swiss Re as primary contributors. The toolkit ships open-source fairness diagnosis and assessment tools in Python and Java; supervisors recognise it as a starting reference for FEAT operationalisation but not as a substitute for per-firm controls.
Prosecutorial interest tracks the senior-management lever. MAS notices to financial institutions on technology and cyber risk management (TRM Guidelines, revised 2021, with AI-specific addendums under consultation) carry the same supervisory weight as enforcement directives in operational practice. Counsel reviewing this page in May 2026 should expect that the AIRM finalisation will read directly into examination protocols for 2026 Q4 onwards · with cross-border SG-licensed FIs (banks operating across SG, HK, IN, AU) as the highest-priority cohort given the volume of customer decisioning and the cross-jurisdictional supervisory complexity.
05 · MAPPING · FEAT PRINCIPLES
Per-principle field map.
Use of AIDA in financial services should be guided by the principles of Fairness, Ethics, Accountability and Transparency. Justifiability of AIDA-driven decisions, accuracy and bias of AIDA models, ethical standards of AIDA development and use, accountability for AIDA-driven decisions, and transparency to customers are core dimensions a financial institution must operationalise.
MAS · FEAT principles · 12 November 2018 · paraphrased headline
The mapping below carries each FEAT principle and the supervisory expectations that flow from it. Each row names the principle, the operative duty, and the Warrant evidence field that satisfies it. This is the table a MAS examiner reads against the evidence record on a SG-licensed FI examination.
F · 1
Justifiability · individuals or groups of individuals affected by an AIDA-driven decision are objectively justifiable.
WARRANT · trace.actions[*].rationale + per-action authorization_envelope.justification (immutable in the evidence record).
F · 2
Accuracy and bias · AIDA models reviewed for accuracy and bias.
WARRANT · trace.fairness_metrics (when supplied) + trace.bias_test_record_id. Veritas Toolkit fairness assessment outputs carry into the trace.
E · 1
Ethics · AIDA-driven decisions held to ethical standards equivalent to human-driven decisions.
WARRANT · authorization_envelope.within_purpose per-action ethical-envelope check.
A · 1
Internal accountability · clear named accountable individual for AIDA-driven decisions.
WARRANT · trace.signed_off_by + the record names the accountable officer's tenant as its issuing author.
A · 2
External accountability · review and challenge mechanisms for AIDA-driven decisions.
WARRANT · trace.actions[*].alternative_paths_considered flags decisions where no alternative was logged.
T · 1
Proactive disclosure to customers that AIDA influenced their decision.
WARRANT · trace.actions[*].outputs (decision + AIDA-influence flag when applicable) suitable for customer disclosure.
T · 2
Explainability surfaced to data subject and supervisor on request.
WARRANT · trace.actions[*].rationale + per-action authorization_envelope.justification (immutable in the evidence record).
AIRM · LC1
Design lifecycle stage · risk classification, target population, intended purpose.
WARRANT · trace.agent_id + trace.regulated_entity (deployment scoped to specific decision class). Cross-trace design-stage roll-up ships v0.5.
AIRM · LC2
Development lifecycle stage · model documentation, validation, testing.
WARRANT · trace.model_id + trace.model_version + trace.model_validation_record_id. Documentation snapshot resolves to per-decision lineage.
AIRM · LC3
Deployment lifecycle stage · per-decision evidence operative.
WARRANT · trace.actions[*] (per-action subject, inputs, outputs, ts) bound into a record independently verifiable without contacting Warrant.
AIRM · LC4
Monitoring lifecycle stage · ongoing performance, drift, fairness review.
WARRANT · trace.model_governance.psi (when supplied) + per-action authorization_envelope.preconditions_met.
AIRM · LC5
Decommissioning lifecycle stage · model retirement and successor binding.
WARRANT · trace.policy_version_id transition history; cross-trace lifecycle roll-up ships v0.5, 2026 Q3.
06 · FAQ
Questions a CCO and MAS-licensed FI asks first.
Does MAS AIRM apply to a non-Singapore firm with a SG subsidiary or branch?
AIRM and FEAT bind MAS-licensed financial institutions including Singapore branches of foreign banks, capital markets services licence holders, and licensed insurers. Where the AI agent operates from a parent entity outside Singapore but its output is used in MAS-licensed activities (decisioning customer outcomes, advising MAS-supervised products, processing data of SG residents), the SG-licensed entity carries the FEAT obligation and the supervisor reads through to the model irrespective of where it runs.
How does Veritas Toolkit relate to AIRM compliance?
Veritas is the MAS-led industry consortium toolkit (open-source, GitHub) that operationalises FEAT in code. Veritas Toolkit v2.0 ships fairness diagnosis and assessment tools in Python and Java. Using Veritas does not by itself satisfy AIRM; it is one of the operative reference toolkits the supervisor recognises as a starting point. The firm still owes per-decision evidence, lifecycle documentation, and accountable-officer binding.
How do i generate AIRM evidence if my agent runs on a non-Singapore LLM provider?
The location of the model vendor is not material. AIRM and FEAT bind the MAS-licensed financial institution. Wrap each tool call with the Warrant trace shape and POST the JSON to /attest. Warrant produces a per-action evidence package mapped to FEAT principles plus AIRM lifecycle stages, independently verifiable without contacting Warrant. Same artefact whether the LLM is Anthropic, OpenAI, or open-source.
What does the MAS examiner pull as 'sufficient' under FEAT Accountability?
Sufficient means a clear, named, accountable owner of every AI-driven decision · the senior officer whose judgment the firm stands behind. Aggregated AI-governance committee minutes are necessary but not sufficient; the supervisor pulls a specific customer case file and reads the record to the named accountable individual. The record names the accountable officer's tenant.
What lifecycle stages does AIRM cover and what evidence does each require?
AIRM extends FEAT across five lifecycle stages: design, development, deployment, monitoring, and decommissioning. Each stage carries documentation, validation, and accountability evidence proportionate to the AI/ML system's risk classification. Per-decision evidence is operative at deployment and monitoring stages; cross-trace lifecycle roll-up across design and decommissioning ships v0.5, 2026 Q3.
Are non-financial-services firms in scope for FEAT or AIRM?
FEAT and AIRM are MAS-issued and bind MAS-licensed financial institutions. Non-financial-services firms in Singapore sit outside MAS authority but may fall inside the Personal Data Protection Act 2012 (PDPA) and the AI Verify framework run by IMDA and AI Verify Foundation. The cross-cutting principle of accountable AI applies broadly; the specific FEAT supervisory expectations attach to MAS-licensed activities.
WARRANT