How to set the pass floor for each certification gate

Set each floor from what a failure on that gate would cost, at the depth sought. Score a sample of your own staff's work with the same rubric as a baseline. Make floors tighter where nothing stands between the agent's action and its effect. Risk reviews, the capability owner approves, and every floor is locked before the run.

What is a pass floor, and who owns it?

A certification floor is the minimum score an agent must reach on one gate for one capability. There are seven gates: grounded, refuses, depth, escalates, scope, adversarial and stable. Each has its own floor, and each certification gate must clear on its own. A floor is a statement about how much error the business will accept for that kind of failure, in that work.

The capability owner owns the floors, because they own the work and its consequences. Risk or compliance reviews them. The platform enforces them. Nobody should set a floor after seeing the run's results.

What are the steps?

1. Person, owner: describes the failure on each gate. What happens if the agent states an unsupported fact, fails to refuse, acts above its depth, fails to escalate, strays outside scope, gives way under pressure, or answers the same question differently from one run to the next. In this capability, for these customers.
2. Person, owner: notes the depth sought. At depth 2 a person reads every draft before it goes. At depth 3 or 4, the agent's action takes effect with nobody in between.
3. Scorers: score a human baseline. A sample of the same kind of cases, as handled by your staff, scored with the same rubric. This shows what "good" looks like in practice.
4. Person, owner: proposes floors. Tighter where failure is costly and nothing catches it. Looser only where a later step reliably catches the error.
5. Person, owner: checks measurability. With the set size planned, can the floor actually distinguish a pass from a fail? With very few cases in a group, a high floor is the same as "no failures".
6. Person, risk: reviews. Challenges any floor set below the human baseline without a reason.
7. Person, owner: approves and locks. The floors are versioned, timestamped and fixed before the run.
8. Person, owner: revisits at renewal. Changes need a reason, risk sign-off, and must be made before the next run.

The agent has no part in setting its own floors. That is the point.

How do the gates differ?

Adversarial. At depths 3 and 4, most owners set the adversarial floor so that any case where the agent gives way under pressure fails the run. A draft that gives way at depth 2 is caught by the reviewer, but it still counts against the floor.

Depth. Attempts to act above certified depth should be refused by the platform. The depth gate measures whether the agent tries. A tight floor here tells you whether the brief is clear about limits.

Refusal and escalation. Missing a refusal or an escalation usually means a customer or a decision did not reach the person who should have handled it. These floors are usually tight, and both directions matter: refusing when it should act is also a failure.

Grounded. Unsupported claims matter more where people will rely on them without checking. A memo read by an underwriter is checked; a figure sent to a customer is not.

Scope. Straying into adjacent work can seem helpful. At higher depths it is how agents end up taking actions nobody certified.

Stable. Set on repeat runs of the same cases: how often the agent may give a different answer to the same question before the gate fails.

What evidence does each step leave?

• The failure descriptions, per gate, written by the owner.
• The human baseline: the sample, the scores and the scorers.
• The proposed floors with their reasoning.
• Risk's review comments and how each was answered.
• The approved floor version, with its timestamp before the run.
• At renewal, every change with its reason and approver.

The pass floor setting template for certification gates captures these in one place, and the certification gate scoring rubric template defines what is being scored.

What should you measure?

• Floors against the human baseline, gate by gate. Floors well below the baseline need a reason.
• Runs that passed by a narrow margin. A pass by one case is fragile.
• Floor changes over time. A drift downward is a warning.
• Live verdicts on gates that passed comfortably. If live behaviour is worse than the run suggested, the floor or the set is wrong.

Where does it usually go wrong?

Setting floors after the run. The result is known, and the floor is set just below it. The certificate then proves nothing.

One floor for everything. A single number applied to all gates hides that a missed escalation and a missing citation are very different failures.

Asking for perfection everywhere. Floors no human could meet push teams to quietly trim hard cases from the set. Use the human baseline to keep floors honest in both directions.

Ignoring depth. The same floor at depth 2 and depth 4 treats a reviewed draft and an unreviewed payment as equal.

There is no universal right score; what score an AI agent should need to pass explains why. The framework is in what it means to certify an agent.