What does 5.17 require?

The development of AI is written up in enough depth to be reproduced and audited. How much depth depends on the materiality of the use case, but the record must let an outsider, a reviewer or an auditor, understand and in principle rebuild the implementation of the system or model and rerun its evaluation and testing. Documentation standards span the whole development effort and may take in data sources, processing and quality checks; why the chosen approach was selected; training procedures with code versions, environments and hyperparameters; evaluation measures, thresholds, test methods and results; the explainability analysis and fairness assessment; and the main assumptions, limitations and mitigants.

Singapore MAS Guidelines on Artificial Intelligence Risk...

Other controls in Section 5 AI life cycle controls: data, transparency, fairness, human oversight, third-party AI, selection, evaluation and testing, technology and cyber, reproducibility, pre-deployment review, monitoring, change management and retirement (paragraphs 5.1 to 5.26) – Singapore MAS Guidelines on Artificial Intelligence Risk Management (2026)

5.1-5.2 Plan and implement controls across the entire AI life cycle with clear roles, reviewed regularly for newer technologies, calibrated to risk materiality, with constraints such as limited third-party disclosure mitigated (including by limiting AI use) so that residual risk stays within appetite, and partial deployments governed by explicit policies
5.10-5.11 Third-party AI: onboarding, development and deployment controls adequate to materiality, contractual visibility over AI introduction and changes, testing in the FI's own context with compensatory testing for disclosure gaps and documented suitability; the FI retains accountability and weighs transparency, supply chain, concentration, change notification, contingency, legal terms, staff capability and complexity, limiting, suspending or replacing the service where residual risk exceeds appetite
5.12-5.13 Select AI systems, models and features against the objectives and risks of the use case, documenting the trade-offs between performance, complexity, fairness and explainability in proportion to materiality, with domain expert or user review, and weigh the heightened risks of newer, less understood systems against the FI's ability to mitigate them
5.14-5.15 Evaluate and test each system or model before deployment to reliability and safety levels proportionate to materiality: identify key risks and set measurable thresholds, define and agree evaluation measures, use testing approaches such as out-of-sample and out-of-time testing, sensitivity, stability, sub-population, stress and adversarial testing, error analysis and benchmarking on representative data, mitigate overfitting, and put guardrails in place for identified limitations, covering the failure modes of Generative AI and agents
5.16 Technology and cybersecurity controls for AI systems commensurate with materiality under MAS technology risk requirements: secured environments with hardened configurations, segmentation, input validation, API authentication, encryption and data loss prevention; tightly controlled role-based and multi-factor access with privileged access management and separation of duties; and governance, data exposure restriction and monitoring of third-party components such as plugins and APIs
5.18-5.21 Pre-deployment reviews by parties not involved in development, under clear policies with assigned roles and scope and independence proportionate to materiality; formal independent validation by competent, objective validators for high-materiality use cases covering conceptual soundness, data, implementation integrity, evaluation results, explainability and fairness, and assumptions; documented peer review for others; findings, limitations, remediation and conditions reported to the approval body, which ensures they are actioned
5.22 Technology and cybersecurity reviews before deployment covering technical implementation, system and network security, resilience, recoverability and operational readiness: secure design and least-privilege access review of architecture, APIs, plugins and third-party services with go-live checklists confirming encryption, data loss prevention, firewalls, access restrictions and logging; vulnerability assessment, penetration testing and red teaming, adversarial testing and safeguards such as input validation, throttling and anomaly detection
5.23 Ongoing monitoring of all deployed AI including third-party AI, proportionate to materiality: defined metrics and tiered thresholds for robustness, stability, data quality and fairness with data, concept and model drift checks and monitoring of reasoning, actions and tools in AI workflows; incident and issue reporting, escalation and resolution including retraining, redevelopment or decommissioning, kill switches for high-materiality AI and user feedback channels; accountable roles; records of monitoring and remediation with controlled, logged access to models, data, pipelines and configuration and enhanced logging for Generative AI and agents; and training for monitors and users

Query this from an agent

The graph holds this control, the 0 it maps to, and the evidence behind each claim, over MCP and REST.

MCP endpoint All Singapore MAS Guidelines on Artificial Intelligence Risk Management (2026) controls