1. Purpose
Openwist builds AI systems, from language models and assistants to software for vehicles and robots. This framework sets out the principles we follow and the requirements we apply to decide whether, when and how a system is released. It is written so that its commitments grow stricter as the potential for harm grows, and so that it can stay in place as our products change.
It is a statement of how we operate. It is not a warranty, and it does not replace our Terms of Service or Usage Policy.
2. Principles
- People first. Our systems exist to serve people. Humans must be able to understand, question, correct and override what they do.
- Proportionate caution. The more harm a system could cause, the more evidence, oversight and delay we require before release.
- Evidence before claims. We state only what we have tested, and we say how it was tested.
- Design for understanding. We prefer architectures and methods whose behaviour we can explain, test and reproduce.
- Defence in depth. No single safeguard is relied on alone. Safeguards sit in the model, the software around it, the product, and the operating environment.
- Traceability. We can say where our data came from, what versions of a system were released, who approved them, and what changed.
- Accountability. A named role owns each release and each incident.
- Learn in the open. We publish what we learn where it is safe to, including failures, and we invite scrutiny.
3. Scope and terms
This framework covers every model, product, service and piece of software that Openwist develops or releases, including its research code, public tools, Exa, and software for vehicles and robots. It also covers the data used to build and evaluate them.
A release is any point at which a system becomes available outside the team that built it, or begins to affect real people or real hardware. A safety case is a written, structured argument, backed by evidence, that a system is acceptably safe for a stated use in a stated environment.
4. Governance
- Accountable executive. The Chief Executive is accountable for this framework and for every release decision, and may delegate review to a designated safety lead.
- Release gates. Each release is reviewed against the requirements of its level (section 5) and is approved in writing by someone who was not the main author of the system, wherever the size of the organisation allows. For Level 4 systems, the reviewer must be independent of the team that built the system.
- Records. We keep a decision register that records major decisions, the options considered, the evidence, the reasoning, who decided, and when. Open decisions are tracked until they are closed.
- Authority to stop. Any person at Openwist may raise a safety concern, and a designated safety lead can pause a release until it is resolved.
- Review. We review this framework at least once a year and after any serious incident, and we version every change.
5. Risk levels and required safeguards
We classify each system by what it can do and what could go wrong, and apply the requirements for the highest level it reaches.
| Level | Description | Examples |
|---|---|---|
| 1. Informational | Produces text, images, audio or analysis for a person to use | Language models, chat assistant, developer tools |
| 2. Agentic software | Takes actions through tools, code execution, files, accounts or messages | Assistants that run code, browse, send messages or create files |
| 3. Physical-system software, not in control | Software for vehicles and robots that runs in simulation, on test benches, or in an advisory role | Planners, controllers and perception stacks in simulation or on a bench |
| 4. Physical-system control | Software that operates a real vehicle, robot or other machine | Driving and robot control software on hardware |
| Requirement | L1 | L2 | L3 | L4 |
|---|---|---|---|---|
| Documented purpose, intended and unsupported uses | ✓ | ✓ | ✓ | ✓ |
| Pre-release evaluation for harmful output, accuracy, privacy and robustness | ✓ | ✓ | ✓ | ✓ |
| Usage Policy and enforcement | ✓ | ✓ | ✓ | ✓ |
| Monitoring for misuse and failures after release | ✓ | ✓ | ✓ | ✓ |
| Isolation of execution, least-privilege permissions, and confirmation before irreversible actions | ✓ | ✓ | ✓ | |
| Testing against prompt injection and tool misuse | ✓ | |||
| Hazard analysis and a safety case | ✓ | ✓ | ||
| Simulation and hardware-in-the-loop testing with recorded results | ✓ | ✓ | ||
| Staged real-world testing with independent hardware stop | ✓ | |||
| Independent review and, before wider deployment, external assessment | ✓ | |||
| Event logging sufficient to reconstruct incidents | ✓ | ✓ | ✓ |
6. Evidence standards
Every claim in our documents, results and marketing carries one of these evidence states, and we do not describe a lower state as a higher one.
| State | Meaning |
|---|---|
| Assumed | A value or fact taken from a source or an estimate, not yet checked by us |
| Illustrative | A worked example showing how something would behave, not a result |
| Simulated | A result from a model or simulation, with the software version, settings and seeds recorded |
| Verified | A result reproduced independently, for example on a second machine, a second environment, or by another party |
| Validated | A result confirmed against real-world measurements or operation |
Further rules:
- Results can be reproduced from recorded code, data and settings, and we run important results in a clean environment before relying on them.
- Provisional values, such as unconfirmed vehicle parameters, are labelled as provisional wherever they appear and are not hard-coded into designs.
- We report negative and mixed results alongside positive ones.
- Timing and performance measurements are reported as measured, and we do not claim guarantees we have not demonstrated, for example worst-case execution time.
- Data carries its source and licence, and personal data is handled under our Privacy Policy.
- We separate what a design requires from what it achieves, and we keep a register of open decisions.
7. Safeguards for AI models and Exa
Before release we evaluate each model and assistant for: harmful output across the categories in the Usage Policy; accuracy and invented information; privacy leakage and exposure of other users' data; manipulation, jailbreaks and prompt injection; behaviour under unusual or adversarial inputs; and bias in significant uses. We repeat these evaluations on each significant new version and record the results.
In operation we provide and maintain:
- the Usage Policy and enforcement, including warnings, restrictions and account action;
- automated and human review to detect abuse, with appeal through support@openwist.com;
- limits that apply by plan, to reduce abuse and protect capacity;
- clear notices that AI output can be wrong and should be checked;
- isolated environments for code execution, with restricted network and file access;
- confirmation steps before actions with real-world effects;
- reporting channels for harmful output at safety@openwist.com;
- user choice over whether conversations may be used to improve our models; and
- protections for personal data under our Privacy Policy; and
- a minimum age of 13, parental permission for users under 18, additional safeguards for younger users, and no use of their conversations to train our models.
Models released openly. Before we release model weights or tools publicly, we consider whether they would give serious uplift to someone seeking to cause large-scale harm, and we withhold or limit what would. We document the model's purpose, training data summary, known limitations and intended use in a model card.
8. Physical systems: vehicles and robots
Software that moves machines can injure people in ways that text cannot, so Level 3 and Level 4 systems follow additional requirements.
Design and analysis
- Identify hazards and hazardous events, rate their severity, and trace each to a safety requirement, using recognised methods such as hazard analysis and risk assessment and system-theoretic analysis.
- Define the operating design domain, meaning the conditions under which the system is allowed to operate, and make the system refuse or hand over outside it.
- Provide defined safe states and minimal-risk behaviours for faults, loss of sensors, loss of communication and loss of power.
- Add independent runtime monitors that can override the primary software and bring the machine to a safe state.
- Include a physical emergency stop and an independent hardware safety path that does not rely on our software.
- Protect the system against cyber attack, including secured communications, authenticated and signed software updates, and controlled over-the-air changes.
Verification, in order
- Unit, component and integration tests, including fault injection.
- Large-scale simulation across normal, rare and failure conditions, including adverse surfaces, weather and sensor degradation, with results recorded and reproducible.
- Hardware-in-the-loop and bench testing.
- Closed-area testing with a trained human able to stop the machine immediately.
- Supervised limited use, with restricted speed, force, range and location, and a human supervisor.
- Wider deployment only when the earlier stages have been completed with acceptable results, an independent review has approved it, and external assessment has been obtained.
We move forward only when exit criteria for each stage are met and recorded, and we step back if they stop being met.
Standards. We use relevant recognised standards as reference points and align with them where they apply to the product, including functional safety, safety of the intended functionality, and cybersecurity standards for road vehicles, safety standards for robots, and the regulations of the places where we operate.
Human oversight. No Openwist system will operate a vehicle or robot with people nearby unless a human can take control of it or stop it. Any capability that reduces that oversight requires a new safety case and a new release decision.
Records. Level 3 and Level 4 systems log the data needed to reconstruct what the system sensed, decided and did around any incident, handled under our Privacy Policy.
Constitution. The Openwist Constitution for Vehicles and Robots sets out what our vehicles and robots will and will not do. It forms part of this framework.
9. Security
We apply secure development practices, review dependencies, manage secrets, control access, and monitor our systems. We accept vulnerability reports under our Responsible Disclosure Policy and treat flaws that could affect physical safety as the highest priority.
10. Incidents
An incident is any event in which a system causes or nearly causes harm, exposes personal data, or fails a safeguard in a significant way.
- Contain. Limit the effect, including pausing or withdrawing the affected system or feature.
- Assess. Rate the severity and identify who is affected.
- Notify. Tell affected people and the regulators and partners the law or agreements require, within the time allowed.
- Fix. Find the root cause, correct it, and add tests so that it is caught in future.
- Review. Hold a written post-incident review with named owners and dates for each action.
- Share. Publish a summary of significant incidents and what changed, unless that would increase risk or breach someone's privacy.
11. Transparency and external scrutiny
- We publish model cards and system documentation for what we release, and papers for research where it is appropriate to do so.
- We publish updates on this framework, the changes made, and the reasons.
- We welcome independent researchers and report-finders under our Responsible Disclosure Policy.
- Before a Level 4 system moves beyond supervised testing, we seek an independent assessment, and we publish a summary where possible.
12. Changes
We will not weaken a commitment in this framework without publishing the change, the reasons, and its date. Each version is kept and numbered.
13. Contact
Safety concerns and questions: safety@openwist.com
Security vulnerabilities: security@openwist.com
Product defects and recalls: recall@openwist.com
Complaints and claims about an incident: accountability@openwist.com