Skip to main content
Safety at Vinci

Use AI with a clearer idea of what to check.

An answer, a shared document, a task you hand over. Each brings a different question about safety. Start with yours, try a small example, and see what our research has taught us.

What would you like to understand?

Can I trust this answer?

Check an important detail against its source. Notice what is supported, what is wrong, and what the source does not say.

Try checking an answer

What happens to my information?

See how running a model on your device, using a hosted service, adding search or syncing a conversation changes where information could go.

Explore the data routes

How do I keep control?

Follow a task from a note to a draft. Decide what may happen next, where approval is needed, and when the task should stop.

Try a task with limits

These short activities use prepared, fictional examples. No account or typing is needed, and no AI runs in them.

What we tried. What we learned.

Useful safety work has to show what happened, including when an approach fell short. These are findings from our published work, with the limits beside them.

Fewer unsupported claims, but fewer useful answers.

What happened
We applied the same training approach to three model families. They made fewer unsupported claims, but also gave fewer correct answers to questions they should have been able to answer. None passed all the study’s checks.
What you can use
When a score improves, ask what got worse and how often the model still gives a correct answer.
Keep in mind
Two other AI models scored answers in an early test. The separate final test was not run, broader abilities were not evaluated, and no trained version was recommended for release.
Read the study and its methods

Our test accepted programs that were wrong.

What happened
An audit found that our original code-checking test accepted all 24 deliberately wrong programs designed to exploit its shortcuts. We withdrew or paused the claims that depended on that test.
What you can use
Ask whether a check can catch plausible mistakes, as well as recognize correct work. A passed check only covers what it actually tested.
Keep in mind
Our replacement test also did not meet its requirements. This study produced no model or set of evaluation tasks for release. These results do not show that checks work reliably for other tasks.
Read the study and its methods

Changing one part of a plan changed another.

What happened
In Conestoga’s recorded Piccolo conversations, a requested change to one day also changed another day’s total. Other replies contained incorrect sums or left the final answer unfinished.
What you can use
Keep the earlier version. After a change, check that untouched details, totals and earlier requirements still hold.
Keep in mind
These observations came from specific conversations, not a test designed to compare model settings. The conversations were not independently repeated. A model rating its own answer is not an independent check.
Read the recorded findings

How this shapes our work

Test before release.

Our published studies show why release decisions need to consider safety and useful answers together. Each model’s information page lists its published tests, results and known gaps.

See models and their results

Make the principles public.

The Vinci Constitution sets out the behaviour we train toward: honesty, useful answers, and refusing harmful requests.

The Constitution describes the behaviour we aim for. Actual answers can fall short; use published tests and results to assess a model.

Current terms and model licences apply.

Read the Constitution

Keep learning after release.

We use community feedback and continued internal testing to record what needs improvement. Reports of harmful behaviour or answers that fall short of our principles guide future model releases.

Existing downloads do not change automatically; improvements need a new release.

Share feedback

Tests reduce uncertainty; they cannot guarantee safe behaviour in every situation. Check each model’s published results and limits before using it.

Technical details and safeguards

Foundry: our internal testing environment

Foundry helps us examine AI tasks, their results and the checks used to assess them. It keeps the context needed to review possible future training examples.

What this does not establish: A defined test environment and a record of checks do not by themselves prove that a result is correct or suitable for training.

Read the evidence

Work, authority, and evidence

Vinci Contracts publishes types and validators for defining AI work, recording evidence, preserving approval boundaries, and expressing the scope of an assessment.

What this does not establish: These are Vinci implementations, not an adopted standard or certification. Schema validation alone does not establish enforcement by a consuming application.

Read the evidence

Execution and authority boundaries

The public Vinci Code demonstrations show approval, review, and explicit completion states. They make some limits visible during AI-assisted technical work.

What this does not establish: Local guards are not a security sandbox. Records of checks do not independently establish correctness; permissions and deployment still need review.

Read the evidence

A responsible-use guide in preparation

We are preparing the Vinci Charter to explain responsible use and the obligations we intend to set for customers. It has not been published; current terms and model licences apply.

Using a model responsibly

Review the actual model, provider, data path, permissions, and human responsibilities for the intended use. Canadian ownership alone does not establish Canadian data residency or the origin of every component. Software pages describe supported modes and their limits.

Published data and policy information

Found something we should test?

Tell us which model or tool you used, what you expected, and what happened instead. Include a public source or a small fictional example if it helps. Leave out personal information, confidential documents and secrets.

Contact the team