Scope
Environments, categories and decision types the qualification covers.
AI Agent Certification
Before your agent represents you, prove its judgment.
Prepare your agents to carry defined work against your organization’s expert standard, and test whether they can do it when the situation changes.
How qualification works
Confidence the expertise transferred to your agent.
01
Scenario
Governed scenarios drawn from your environment: edge cases, boundaries, changed context, the moments that must be refused.
02
Behavior
The agent’s decisions are evaluated against the expert standard, including escalation and stop behavior.
03
Qualification record
Certified or refused, with scope, critical failures, remediation and renewal date.
The deliverable
What the record contains.
Critical failures
Behaviors that disqualify regardless of average performance.
Remediation
What must change before re-evaluation.
Renewal
Revalidation triggers: model, policy, standard or environment change.
Efficiency
A qualified agent applies governed judgment instead of rebuilding it.
Because the standard is formalized upstream, the agent does not reconstruct the decision framework on every call.
Three ways in
Already built an agent? Bring it.
You do not have to commission a new agent to have one qualified, and you do not have to adopt Katya. Most customers arrive with something already running and want to know whether it is safe to widen its scope.
Most common
Your existing agent
Whatever it was built on — your own stack, a vendor platform, an agent framework. It is trained against the standard, tested on your scenarios, and either cleared for named work or refused with the reasons.
You keep the agent. Nothing is rebuilt to qualify it.
Qualify my agent →Built here
An agent built around your work
When there is nothing running yet, the agent is built against your governed standard from the start — then stress-tested and qualified on the same terms as any other.
Scoped as a design-partner delivery.
Build a governed agent →First-party proof
Katya
The first certified Kataclyzim agent, already operating from Resistance Intelligence. Deploy her configured to your environment — or just use her as evidence that the qualification means something.
An option, not a prerequisite.
See Katya's record →Test scope
A fluent answer is not a safe one.
An agent that sounds right on the easy ninety per cent and improvises on the rest is the expensive kind of failure. The assessment is built around the cases where that shows.
The decisions it has to get right
- Reading the situation correctly in the environment it will actually operate in
- Staying inside the offer, the facts and the claims it is permitted to make
- Not inventing context — stakeholders, budgets, approvals or conditions nobody mentioned
- Holding the separation between environments, so one domain's logic does not leak into another
Boundary and escalation behaviour
- Whether it stops when it should, rather than finding a way around the stop
- Whether it escalates at the right point, not three turns after it should have
- Whether the human inheriting the conversation gets enough to continue it
- Whether it refuses cleanly under adversarial pressure instead of degrading
Failure analysis
Where it broke and what pattern the breaks share — not a pass rate. A refusal here is the cheapest one you will ever get.
Applicable work
The work and the environments it is cleared for, and the ones it is not. Never a general-purpose clearance.
Remediation
Targeted changes against the named failures, then a re-test on that scope — your team's work or ours, as scoped.
Revalidation
A model version, a prompt change, a new environment or an incident makes the record stale. The triggers are named up front.
The evaluation method itself is not disclosed. What the evaluation concluded, and what it means for deployment, is. All four certification tracks →