Skip to main content
Nobody switches a production estate from nothing to blocking in one step, so the platform does not ask. Enforcement is a mode on a grant, not a global setting, and each grant moves independently when the team that owns it agrees.
1

observe

The check runs and records what it would have done. Nothing is refused, nothing changes. Zero operational risk, and after a fortnight you know what your agents are actually doing rather than what the design document says.
2

shadow

The full decision path is exercised, including policy evaluation, and still nothing is refused. You can also pass incumbent_decision, which is what your existing system or reviewer decided, and the divergence is the point. Where they disagree is where the policy is wrong.
3

canary

Refusals begin, on a subset. This is the first rung where an agent can actually be stopped.
4

enforce

Refusals apply to everything under this grant.
The first two rungs are free. The commercial line sits where refusals begin to bite, which is the honest place to put it: see pricing.

​
The mode is part of the record

This is the part that matters eighteen months later. The enforcement mode is sealed into each decision at the moment it was made, so an evidence pack states what was enforcing then rather than what is configured now.
A system that reads the mode from current configuration when producing evidence is reconstructing, not recording. Configuration drifts: a limit raised for a migration in March and never lowered is the most ordinary finding there is.

​
Seeing all four at once

The demo runs the same agent at three rungs simultaneously, and the records differ exactly as the ladder claims: the observe and canary copies refunded the same tickets and recorded enforced: false; the enforce copy refused and escalated. That is the argument for shipping this without a flag day, and it is observable rather than asserted.

​
Moving a grant

See Move up the ladder for the call and what to watch while you do it.