A certification gate is one of the separate tests an AI agent must pass, on its own, to be certified to a depth. Attest uses seven: grounded, refuses, depth, escalates, scope, adversarial and stable. Each asks one question of every answer the agent gives, each has its own floor, and clearing six of the seven is a fail.
Why do certification gates matter?
A single blended score hides the failures that matter. An agent can be accurate on nearly every case and still miss the one signal that a customer is in difficulty, or still act beyond its permitted depth once in a while. Averaged together, those misses disappear into a healthy number.
Gates keep each kind of failure visible. An agent that never invents facts but fails to escalate has an escalation problem, and the gate says so. The argument that declining correctly deserves its own test is made in refusal is a gate, not a nice-to-have.
What does each gate ask?
| Gate | The question for each case |
|---|---|
| Grounded | Does every factual statement trace to a source the agent may use? |
| Refuses | Did it decline what it should, and only that? |
| Depth | Did it act no further than its permitted depth? |
| Escalates | Did it hand to a person when a trigger was present? |
| Scope | Did it stay within its role? |
| Adversarial | Did it hold its bounds under pressure? |
| Stable | Did it give the same answer to the same question across runs? |
Each case is scored at every gate that applies, using a rubric written before the run; the certification gate scoring rubric template gives the pass and fail tests. Each gate's score is compared with its own floor, set as described in how to set the pass floor for each certification gate.
What does that look like in an example?
A lettings agency certifies an agent that answers tenants' repair requests. It scores well at grounded, refusal, scope and depth. But in a handful of cases tenants mentioned a smell of gas, and the agent booked a routine repair instead of telling them to leave the property and call the emergency line. Those cases fail at escalation and at adversarial: a smell of gas is a state where the efficient answer is the wrong one. On a blended score the agent looked excellent. Gate by gate, it is not certified until the brief is fixed and the run repeated.
How is a certification gate different from a human gate?
A human gate is a point in a live workflow where a person must approve before an action proceeds, such as a payment held for a manager. It operates at run time, on each action. A certification gate is a test the agent must pass before it is allowed to act at all, operating on a set of past cases. The certification gate decides how much the agent may do alone; the human gate catches what it is not allowed to do alone. You need both, and passing one does not remove the other.
Which related terms should you know?
- Depth gate: tests whether the agent stays within its permitted depth.
- Stable gate: tests whether the agent gives the same answer to the same question across runs and within a conversation.
- Certification floor: the minimum score each gate must reach.
The full approach is in what it means to certify an agent.
Where Pryme fits
Attest scores every answer in a certification run at all seven gates, against floors the customer sets for each gate. An agent is certified only when every gate clears on its own, and the results show which cases failed which gate, so the team can see whether to fix the brief, the sources or the bounds before the next run.
Questions people ask
Why seven gates rather than one score?
Because the failures are different in kind and have different costs. A missed escalation and a wrong citation should not be able to offset each other.
Does every gate apply to every case?
Every gate is scored across the run, but not every case tests every gate. Escalation does not apply where no trigger is present, and stable applies only where the same question is asked more than once, across runs or within a conversation.