Independence
We sell evidence. Nothing else.
We sell no agents, models or fixes and take no referral fees. A pass and a fail pay us the same. If Vidimus is ever acquired, your contract lets you leave with all your data.
- Last updated
- Scope
- Document status
At a glance
What independent AI agent assurance means at Vidimus, in four lines.
- What we sell
- Tests and the evidence they produce. We sell no agents, models or fixes.
- Who owns us
- Founder-owned. No model, cloud or security vendor holds a stake.
- If we are acquired
- You are told, you may terminate, and you have 90 days to retrieve your data. The source code can be placed in escrow on request.
- The one conflict
- The judge is a model, and your agent may run on the same family. Every verdict quotes the reply, so a person can check it.
What we sell and who owns us
- We build nothing we test.
- Vidimus builds no agents and no models, so no verdict can favour a product of ours.
- We sell no fixes.
- A finding never turns into a product we sell. Fixing what we find is your work.
- We take no referral fees.
- We take no commission from any model, cloud or security vendor.
- Founder-owned.
- No model, cloud or security vendor, nor their investment arms, holds a stake. If that changes, this page changes first.
If Vidimus is acquired
The answer is in the contract. Our terms carry a change-of-control clause with three commitments:
You are told.
We notify you of any change of control of Vidimus without undue delay.
You may leave.
You may then terminate and use a 90-day data-retrieval window: your agents, reviews, evidence packs and audit trail stay exportable in full before deletion.
Source-code escrow on request.
If your auditor or a regulator asks for continuity guarantees, we support a source-code escrow, so the software that produces your evidence cannot vanish. Ask during procurement.
The clause is in our terms of use.
How we state and manage the judge’s conflict
One conflict is built in, so we name it. The judge, the model that grades each reply, runs on Mistral, and your agent may too. A judge grading its own model family could, in principle, be lenient. This is what keeps it honest.
| Measure | How it works |
|---|---|
| Every verdict quotes the reply | Every graded verdict carries the words the agent used, so a person can check the grade against the reply. |
| Judge model and instructions on record | Each verdict records the judge model (mistral-large-2512) and the version of its instructions (judge-v7), so it traces to an exact configuration. |
| Calibrated on hand-written cases | The judge gave the expected verdict on 41 of 46 hand-written cases (89%). None of the graded verdicts in that set lacked a quote from the reply. |
| A second opinion on critical failures | A second Mistral model gives its opinion on every failure of a critical clause. It is recorded beside the verdict and never replaces it. |
| A tamper check before the judge | A reply that tries to steer its own grading fails before any model reads it, and the attempt is recorded as a finding. |
| Borderline results go to a person | A Needs review result waits for a reviewer who did not start the run. The database enforces it. |
| Tests reviewed without the author’s reasoning | The model that checks each generated test sees the test and the agent’s profile, never the reasoning of the model that wrote it, so it cannot simply agree with it. |
| A judge from another provider | It is on the roadmap. Until then, every grade can be checked against the reply it quotes. |
Every verdict quotes the reply
- How it works
- Every graded verdict carries the words the agent used, so a person can check the grade against the reply.
Judge model and instructions on record
- How it works
- Each verdict records the judge model (mistral-large-2512) and the version of its instructions (judge-v7), so it traces to an exact configuration.
Calibrated on hand-written cases
- How it works
- The judge gave the expected verdict on 41 of 46 hand-written cases (89%). None of the graded verdicts in that set lacked a quote from the reply.
A second opinion on critical failures
- How it works
- A second Mistral model gives its opinion on every failure of a critical clause. It is recorded beside the verdict and never replaces it.
A tamper check before the judge
- How it works
- A reply that tries to steer its own grading fails before any model reads it, and the attempt is recorded as a finding.
Borderline results go to a person
- How it works
- A Needs review result waits for a reviewer who did not start the run. The database enforces it.
Tests reviewed without the author’s reasoning
- How it works
- The model that checks each generated test sees the test and the agent’s profile, never the reasoning of the model that wrote it, so it cannot simply agree with it.
A judge from another provider
- How it works
- It is on the roadmap. Until then, every grade can be checked against the reply it quotes.
How the judge is kept honest: each measure and how it works.
What the judge sees
The judge reads the clause, the test and the behaviour it expects, the situation the test set, a short profile of what your agent does, and the reply with any tool calls. It is not told which model or vendor runs your agent and never sees its system prompt.
Every step and every threshold is on How we test.
Ask any testing vendor
Ask any testing vendor who owns it, and what it sells besides tests.
Our answers: our founder owns Vidimus, and we sell nothing besides tests and the evidence they produce. Vidimus produces evidence and indicative assessments, pending human review, for your own reviewers to decide on. It is not a certification.
Procurement teams: we put the change-of-control, exit and escrow commitments in your contract. Contact us. How we host and protect your data is on our security page.