Soft launch a block

Turning on a hard block is the moment Sanitized AI starts stopping people’s work. If a guardrail turns out to be noisier than expected, you usually find out through complaints.

Soft launch gives you a window to find out safely: the guardrail warns for two weeks first, then starts blocking on its own.

Background

Soft launch is the bridge between two of the enforcement actions: a guardrail set to Nudge warns but lets people through, and one set to Block stops the prompt entirely. Soft launch schedules the move from the first to the second.

Starting a soft launch

1. Open Guardrails and find the guardrail you want to block.
2. In the Enforcement column, choose Block.
3. A confirmation dialog appears showing what this block would have caught recently — for example, “In the last 30 days, this would have blocked 142 prompts from 37 users.”
4. Choose Start as a nudge for 14 days instead of Enable Block.

The guardrail is now scheduled. It appears in the Scheduled card at the top of the Guardrails page with the date it will promote.

What your team sees during the window

Nothing changes for them yet. The guardrail behaves exactly like a nudge: the warning banner appears, they can sanitize or edit, and they can still click “Send Anyway”. No one is blocked.

Every override is recorded, which is the point of the window.

Note

The banner does not automatically announce the upcoming date. If you want your team to know a block is coming, edit the guardrail and put it in the Warning Banner Note field — for example, “This will be blocked starting August 11.” That note appears directly in the banner.

What to watch for

Open the Events page during the window and look at the Overrides card. It shows what share of flags people sent anyway, and the guardrails overridden most often.

• High override rate on this guardrail — it is probably firing on things that aren’t actually sensitive. Fix the guardrail, or cancel the soft launch, before it starts blocking real work.
• Low override rate — people are sanitizing or editing instead. The block will be uneventful. Let it promote.

What happens on the date

The guardrail promotes itself. There is nothing to run and nothing to confirm.

Sanitized AI stores the change as “this nudges now, and blocks from August 11 onward”, then checks that date every time it scans a prompt. The switch happens for everyone at the same moment, cannot half-apply, and cannot be missed because a server restarted overnight.

Cancelling or promoting early

Both are available at any point during the window, from the Scheduled card on the Guardrails page:

• Cancel — clears the scheduled date. The guardrail stays a nudge indefinitely.
• Promote early — set the guardrail’s Enforcement to Block directly. It takes effect immediately and the schedule is replaced.

Related

• Enforcement actions — what nudge, block, and monitor mean, and how severity picks the default.
• Exceptions — if a specific person or department needs to be exempt from a block, exempt them rather than weakening the guardrail for everyone.
• Detection settings — choosing what counts as sensitive in the first place.