Quality and safety

Guardrail coverage

The eight categories, every sensitivity and action setting on each side, and the readiness checks that watch them.

Reference for what the guardrails cover: the categories, the settings each side offers, and what runs no matter how the page is configured.

The eight categories

Picking an action for the agent's replies does nothing until at least one of the five scored categories is switched on.

CategoryWhat it can do
HarassmentStop a reply
HateStop a reply
SexualStop a reply
ViolenceStop a reply
Illicit or criminalStop a reply
Self-harmSteer the next reply
Child safetySteer the next reply
Regulated professional adviceSteer the next reply

Why the split

The five scored categories map onto content filters that return a confidence, so a match can stop the reply at the sensitivity you chose. The other three have no score behind them, so they are on-and-off switches and they steer the agent's next reply instead. The editor lists the five under Checked before the caller hears it and the three under Conversation guidance.

Sensitivity

Each scored category carries one of four settings, and so does the caller side's Closer check for attempts to trick the agent. The scale describes how sure the check has to be before it fires.

SettingFires whenEffect
OffNeverThe category is not checked
Only when certainThe match is confidentCatches the least, and rarely stops an ordinary turn
BalancedThe match is reasonably likelyThe middle setting
Catch the mostEven a weak matchCatches the most, and will occasionally stop a turn that was fine

Setting all five at once

The How strict control on the agent's card writes all five scored categories together.

SensitivityEach of the five becomes
OffOff
Let most throughOnly when certain
BalancedBalanced
Catch the mostCatch the most

Change one category on its own and the control reads Custom. Custom is a readout of what the five categories say, not a setting you can pick.

Actions

One action covers each side, and each side offers its own list. Watch only records what would have been caught and changes nothing the caller hears. It's a safe starting point for the agent's replies. On the caller's side it also switches the built-in screen to recording only, so see the note below.

When a reply is caught

ActionWhat happens
Watch onlyThe reply is spoken unchanged and the hit is recorded
Replace unsafe replies (default)The reply is cut and the agent speaks the safe line
Put the caller through to a personThe agent speaks the safe line, then transfers the caller. Needs a transfer destination
End the callThe agent speaks the safe line, then the call ends once the line finishes

The safe line is the one you write under Final safe response, up to 500 characters. Left empty, the agent speaks the platform line:

Sorry, I can't help with that. Is there something else I can do for you?

When a caller turn is caught

The editor calls this When the check fires. Two actions here rather than four. The caller's own words are not something the agent can replace or hand to a person, so the only lever on this side is the attempt count.

ActionWhat happens
Watch onlyThe hit is recorded and nothing about the call changes
Count it as an attempt (default)The turn is withheld from the agent and counts one attempt toward the three-attempt limit, and the third attempt ends the call

The graded check and the built-in screen share one count, so three attempts spread across the two still ends the call. On the standard pipeline the action covers the built-in screen too: under Watch only its matches are recorded, not withheld or counted.

What runs regardless of your settings

PieceWhen it runs
The injection screenEvery call and chat message, whether or not guardrails are configured. It cannot be switched off
The three-attempt sign-off and hang-upThe same calls, while the caller action is Count it as an attempt

Both cover real inbound calls and in-app test calls.

Where the checks run

Both sides use the same moderation service, running in Sydney, and there is nothing to configure. A check is capped at two seconds. When a check times out or fails and its action is set to act, nothing unchecked goes through: a caller turn is withheld and the caller is asked to try again, without counting an attempt, and a reply is replaced with the safe line. Under Watch only a failed check changes nothing. See Data residency.

What the validator checks

Three readiness checks watch this page.

Fires whenSeverityMessage
An action is set for the agent's replies and none of the five scored categories is onRisk"The reply action has no content category to check. Enable one, or set When a reply is caught to Watch only. Conversation guidance cannot trigger this action."
The agent is set to speech to speech and a guardrail acts on the caller or the replyError, blocks publishing"These guardrails need to check text before speech or tool use. In Call engine, use separate listening, reply and voice models before publishing."
The agent's replies hand the caller to a person and no transfer action has a number to dialRisk"The guardrails hand the caller to a person, but no transfer action has a number, so the agent reads the safe line and nobody is put through." The message then says which transfer action to add a number to, or to pick another action

The first and third read the action you chose yourself rather than the one the page starts on. Watch only never triggers the first two, because it is not being asked to act on anything.

On this page