GVAI-Gate decides whether an AI agent is safe to put in front of customers, and issues a signed certificate that says so — bound to one agent, one data set, one policy, one date.
You have put an agent in front of customers. It talks to them, commits you to things, and takes their money. Somebody tested it. The result was a percentage, and the percentage was good.
Ninety-five per cent correct sounds excellent, right up until you learn the other five per cent leaked personal data.
Nobody asks how you assured the agent while it is working. They ask afterwards, in writing, and they are rarely in a generous mood.
In each case the acceptable answer is the same, and it is a document: this agent, this data, this policy, this date, signed. “We tested it thoroughly” is not an answer anyone can act on.
Your team almost certainly has tracing already. Conversation logs, token counts, latency charts, a view of every call the agent made. All of it is genuinely useful, and none of it answers the question you will be asked.
Observability tells you what happened. It holds no opinion about whether what happened was acceptable, and it produces nothing you can hand to anybody.
| Tracing and dashboards | GVAI-Gate | |
|---|---|---|
| Written for | The engineers building the agent | The person who has to approve it |
| Produces | A record of what happened | A verdict on whether it may ship |
| Arrives | After it reached a customer | Before release, as a gate that refuses |
| Opinion | None. It shows; you interpret | A named policy, agreed in advance |
| Artefact | A link to a trace, valid while the tool is paid for | A signed document that stands on its own |
| Answers | “What did it do?” | “On what basis did we believe it was safe?” |
The two coexist, and we expect you to keep what you have. Instrumentation is for the team building the thing. GVAI-Gate is evidence for everyone else — the board, the auditor, the customer's procurement team, and the version of you who has to explain the decision in six months.
The order matters. Each step is only possible because of the one before it.
The agent is configured inside GVAI-Gate: prompt, model, knowledge bases, skills, tools. Tools run through servers that belong to the harness, which is what makes the next step possible.
Every tool call is captured server-side — which tool, in what order, with which arguments, and which failed. The evidence is the order existing in the database, not the agent claiming it placed one.
Graded checks are averaged. Gates are counted: a forbidden call, a wrong ordering, a leaked detail, a policy answer given without looking anything up. This distinction is the product.
The configuration is fingerprinted at signing, so the harness can say precisely what has drifted since — a new prompt, a different model, a tool added.
Values are replaced with stable typed tokens before the prompt is sent and restored in the answer afterwards. The same email appearing five times maps to one token — delete the values instead and the agent cannot tell that five mentions are one person, and the reasoning falls apart. The mapping never leaves your infrastructure, and the quality of the redaction is itself measured.
The agent and the cases meet in a single run. What that run records is then split two ways — and only one side can stop a certificate being issued.
Each of these makes GVAI-Gate score your agent lower than a competitor would. That is deliberate. A number you can trust has to be able to disappoint you.
| Position | What it means | What the alternative does |
|---|---|---|
| Gates are counted, never averaged | One leaked card number is a failure, however good the other 199 cases were | Folds it into a 99.5% score that reads as excellent |
| Unmeasured is not a pass | A check that could not run is recorded as not measured — never as clean | Omits it from the summary, which then reads as a pass |
| Unobservable is not zero | An agent whose tool calls cannot be seen is recorded as unmeasured, not as having made none | Silently scores an uninstrumented agent as flawless |
| A certificate states its own limits | Everything outside the tested scope is printed on the certificate itself | Issues a badge that reads as a blanket guarantee |
There is a fifth, and it is the one engineers react to: GVAI-Gate proposes its own fixes and cannot apply them. It reads the failures, proposes changes, then has to prove each one on held-back data the proposal was never shown, after a person has approved it. A harness that edits the system it measures has stopped being a measurement.
Not a seal of approval. A dated, checkable claim about a specific thing that was actually measured — and an explicit statement of the scope it covers.
Every certificate prints what sits outside what was tested. A certificate read as a blanket guarantee is worse than no certificate, because it transfers confidence without transferring evidence. Typically:
It expires, and it never becomes false. What it records stays true forever: on that date, that configuration met that policy over that data. What changes is whether it still describes what is deployed. Change the model, the prompt or a tool, and the certificate is still accurate and no longer relevant. GVAI-Gate tells you the moment that happens.
Signed, dated, bound to one configuration. JSON, HTML and PDF, all derived from the same signed document so the three cannot disagree. Attach it to a security review and the conversation moves on.
Every case, every tool call, every verdict, kept. Not a summary you are asked to trust — the record that produced it, available when somebody wants to look.
Your agent changes weekly. The configuration is fingerprinted, so you are told precisely what drifted and when the certificate stopped describing production.
GVAI-Gate runs inside your own environment, on your own model provider accounts. No customer data reaches us — which also means there is no data-processing agreement to negotiate, and nothing for your security team to decline.
Certification does not have to be a meeting. SDKs let your team run a gate from the pipeline they already use, fail a build on a gate breach, and pull the certificate as part of a release — so assurance moves at the speed the team ships, not the speed of a review cycle.
Three lines, and they do different jobs. The platform is what your team works in. The certificate is the thing you hand to someone else. The services are us doing the work alongside you when that is faster than doing it yourself.
Most organisations start with one agent already in production, certify it, and use that first certificate to decide how far to take the rest.
Bring us one agent already talking to customers. We will tell you what it actually does — and whether it passes.
A real conversation, not a form. Tell us what your agent does and we will tell you what we would test.