Model Governance Policy

Last Updated: 28 September 2026

Contact: info@myntriq.io

Company: Myntriq Pte Ltd (UEN 202537571M), Singapore

Purpose

This policy describes how Myntriq selects, routes, evaluates, and governs the AI models available through MyntriqOS. It is intended for customers, procurement teams, and technical reviewers who need to understand how model decisions are made and what controls exist over AI model behaviour on the platform.

This policy names the models and explains the governance. The per-provider facts that change over time — the jurisdiction where inference is processed, whether the provider may use customer content to train its models, and the provider's retention posture as reviewed on a stated date — are recorded in the Myntriq Subprocessor List, which is the authoritative annex to this policy. The clause-level evidence behind those entries is maintained in Myntriq's model provider data-policy review (September 2026).

1. Model Architecture

MyntriqOS uses a multi-model architecture. Rather than being tied to a single AI provider, MyntriqOS routes inference requests through a managed model routing layer that can serve multiple models from multiple providers.

This architecture provides:

• Provider independence — if a single provider experiences an outage or changes its terms, the platform continues to operate through alternative models
• Model choice — customers can select models that match their requirements for capability, cost, or data handling
• Future flexibility — new models can be added to the routing layer as they become available, without requiring customers to change how they interact with the platform

All model inference is handled server-side through Myntriq's managed model routing service. Customers and users interact with a unified platform interface — they do not interact directly with model provider APIs.

One model runs on Myntriq's own infrastructure rather than a third-party provider: the bge-m3 embedding model, which supports knowledge base retrieval. Self-hosted inference is the pattern Myntriq is extending to further models over time (see Section 4).

2. Supported Models

The following models are currently available through MyntriqOS:

For every provider named above, the jurisdiction where inference is processed, whether customer content may be used to train the provider's models, and the provider's retention posture as reviewed on 27 September 2026 are recorded in the Subprocessor List. This policy cross-references that annex rather than restating those details, because provider terms change and the annex carries dated snapshots.

3. Model Selection Principles

Myntriq evaluates models against the following criteria before making them available on the platform:

Capability — The model must be capable of performing the tasks for which it will be used in MyntriqOS (reasoning, instruction following, structured output generation, content creation). Capability is assessed through internal testing against MyntriqOS use cases.

Safety and alignment — The model must have demonstrated reasonable alignment with safe and responsible AI practices, including resistance to generating harmful content and responsiveness to safety instructions.

Data handling — The model provider must offer clear terms around how data submitted through their API is used. Myntriq requires that the provider's terms confirm customer content is not used to train the provider's models. As of 27 September 2026 this requirement is met by every provider on the platform with one transitional exception: the Moonshot AI route (kimi-k3) operates today under platform terms that permit customer content to be used for model improvement, with no opt-out. Myntriq is transitioning this route to Moonshot AI's international platform (Moonshot AI PTE. LTD., Singapore) under a written agreement restricting content use; this exception will be removed when that written restriction is in force. While it is in effect, Myntriq limits this route to lower-sensitivity workloads.

Reliability — The model must be reliably available through a hosted API or proxy service, or run on Myntriq's own infrastructure. Myntriq does not offer models that require customer-managed hosting.

Commercial viability — The model must be available at a cost level that is consistent with the platform's pricing structure and the expected usage patterns of MyntriqOS customers.

4. Open-Source and Open-Weight Models

MyntriqOS includes support for open-source and open-weight AI models — currently Llama 3.3 (Meta), GLM-5.2 (Zhipu), and Qwen 3.7 Plus (Alibaba), all reached through OpenRouter, plus the self-hosted bge-m3 embedding model. Open-source model inclusion reflects several important values:

Independence from proprietary models. Open-source models are not controlled by a single commercial entity. This reduces the risk that changes to a proprietary model provider's terms, pricing, or capabilities create a problem for customers without alternatives.

Cost efficiency. Open-source models are generally available at lower per-token costs, making them suitable for high-volume, lower-complexity tasks (routine query handling, summarisation, content drafting).

Transparency. For customers in certain regulated sectors or with specific procurement requirements, open-source models may be preferred because the model architecture is publicly documented.

Myntriq currently routes hosted open-source and open-weight inference through OpenRouter, which provides hosted inference without requiring Myntriq or its customers to manage model infrastructure. OpenRouter itself does not use inputs or outputs for model training and does not store prompts or completions by default. OpenRouter passes each request to a downstream hosting provider; Myntriq configures its OpenRouter routes to exclude downstream providers that permit training on customer content and to prefer zero-retention endpoints. Inference through OpenRouter is processed in the United States, and no Singapore routing option exists — the full details are recorded in the Subprocessor List.

One OpenRouter route, qwen3.7-plus, terminates at an Alibaba-hosted endpoint. Whether the terminating platform is Alibaba's international or mainland-China platform could not be confirmed from the provider's documentation as of 27 September 2026; this is recorded in Myntriq's provider policy review. Customers who require certainty on this point can disable the route for their organisation through model permissions (Section 5) or contact Myntriq.

The bge-m3 embedding model runs on Myntriq's own infrastructure inside its own cloud environment. No customer data leaves Myntriq's environment for embedding workloads, and no third party processes it. Self-hosted Qwen inference is planned on the same pattern; when it enters service, the corresponding hosted Alibaba Cloud route will be removed from the data path and from the Subprocessor List.

5. Model Permissions and Customer Control

Organisation-level model permissions

Each MyntriqOS customer organisation has a model permissions configuration that determines which models are available within that organisation. This allows:

• Customers to restrict agents to specific models that meet their internal data handling or compliance requirements
• Customers to enable or disable specific models for their users — including disabling any route whose processing jurisdiction or training posture does not meet the organisation's requirements
• Administrators to configure model permissions through the platform settings

Default model permissions are configured during environment setup and can be adjusted by organisation administrators.

Agent-level model selection

Where agent-level model selection is supported, administrators can configure specific agents to use specific models. This allows a customer to run different agents on different models — for example, using a more capable model for executive intelligence tasks and a cost-efficient model for routine query handling.

User model selection

In standard deployments, users interact with agents — they do not select models directly. Model selection is an administrative function. This ensures that model governance decisions are made at the organisation level, not by individual users.

6. Governance Controls

Audit logging

Every AI model invocation through MyntriqOS is recorded in the audit log with:

• The model used
• The agent identifier
• The user identifier
• Timestamp
• Token usage (input and output)
• Estimated cost
• Outcome

This provides administrators with complete visibility over which models are being used, by whom, for what, and at what cost.

Cost visibility

The Governance Dashboard provides cost visibility at the organisation, agent, and user level. Cost data is derived from token usage metrics and the current pricing of each model. This allows organisations to monitor AI spend and to make informed decisions about model selection and usage policies.

Usage limits

Myntriq reserves the right to apply usage limits at the organisation or user level to prevent abuse or to manage platform capacity. Where usage limits are applied, affected customers will be notified.

7. Adding New Models

Myntriq evaluates new models on an ongoing basis. The criteria in Section 3 apply to all new model additions.

When a new model is added to MyntriqOS:

• The Subprocessor List is updated to include the model provider if not already listed, with its processing jurisdiction, training posture, and a dated retention snapshot
• The MyntriqOS Privacy Policy and this Model Governance Policy are updated to reflect the new model
• Customers are notified through the platform or by email at least 14 days before the model becomes available in production
• New models are initially added to the available models list but not activated by default in existing organisations — administrators must opt in

8. Removing or Deprecating Models

If a model is removed from MyntriqOS — due to provider changes, performance issues, security concerns, or strategic reasons — affected customers will be notified with reasonable notice (minimum 14 days where feasible) and provided guidance on migrating agents to alternative models.

9. Model Limitations

Customers should be aware that:

• All AI models, regardless of provider, may produce incorrect, incomplete, or biased outputs
• Model capabilities change over time as providers update their models — Myntriq cannot guarantee that a specific model version will be available indefinitely
• Model behaviour may vary based on the prompt, context, and system instructions provided
• No model is certified for use in regulated decision-making without human review

The AI Usage Policy contains further guidance on the limitations of AI-generated outputs and the customer's responsibility for reviewing them.

10. Contact

For questions about model governance, model availability, or to request information about specific model data handling terms, contact info@myntriq.io.