# Built for trust. Designed for what’s next.

> The principles behind Synainesi: policy set centrally and enforced on the device, private by default, approvals that bind, evidence without surveillance.

Canonical: https://synainesi.com/trust/

Principles

## Trust is designed in, not added on.

A rule that only lives in an interface is not a control. Synainesi is built so that your rules hold where the work actually happens.

### Enforced where the work happens

Policy is set centrally and applied on the device, so a disabled model stays disabled.

### Private by default

Confidential work stays on the device. The cloud governs; it does not collect.

### Model-independent

Use the models your organisation approves, and change them without changing your rules.

### Approvals that bind

When a person approves an action, the approval covers exactly that action, once.

### Evidence without surveillance

Oversight rests on records of decisions, not on copies of everyone’s prompts.

### Honest about data

Masked data is described for what it is. Precision is part of trust.

Open by design

## The controls that govern AI should be open to see.

Our vision is an open core. The heart of Synainesi, the layer that decides what may leave a device and records why, open source for anyone to inspect, challenge and improve.

Trust in AI governance should not depend on taking a vendor’s word for it. It should come from being able to see how the rules are kept.

Open to inspectFree from lock-inBuilt with the community

The core principle

## The cloud defines policy. The device enforces it.

Further reading: [A setting is not a control](https://synainesi.com/insights/a-setting-is-not-a-control/) · [Evidence without surveillance](https://synainesi.com/insights/evidence-without-surveillance/) · [Pseudonymisation is not anonymisation](https://synainesi.com/insights/pseudonymisation-is-not-anonymisation/)

Concept design shown with sample data.

---

Synainesi: AI within your rules.

Source: https://synainesi.com/trust/. The HTML page is canonical; this Markdown version is provided as a convenience.