A standing instruction to withdraw authority when a behavioural contract breaks. Ships observing, and arming it requires stating what it would have done to real history.
A breach of a behavioural contract is a
finding. A circuit breaker turns a finding into an action, so authority is
withdrawn without waiting for somebody to read a screen.This is not request rate limiting. That is infrastructure, and your reverse
proxy already does it. This is a control that revokes authority on the
strength of what the record now says.
The property that makes a breaker safe is not tuning. It is that breached
means the 95% confidence interval lies entirely past the limit — see
conformance claims — and a breaker only ever
reads breached.So a breaker cannot trip on an agent at 5.4% against a 5% ceiling over 40
decisions. That window is not_demonstrated, and nothing in the breaker path
looks at not_demonstrated.
A breaker that revokes authority from a working agent is a worse outcome than
the breach it was meant to prevent, and it is how a control gets uninstalled.
The statistics close that case by construction rather than by a threshold
somebody has to get right.
A breaker ships in observe and records what it would have done.Arming it requires stating how many times it would have tripped over the last 90
days — and the number has to match what the platform computes:
# What would this have done?curl https://your-deployment/v1/agents/$AGENT_ID/breakers/counterfactual \ -H "x-api-key: $ROTASCALE_API_KEY"# Arm it, stating what you sawcurl -X POST https://your-deployment/v1/breakers/$BREAKER_ID/arm \ -H "x-api-key: $ROTASCALE_API_KEY" \ -H "content-type: application/json" \ -d '{"basis": "Agreed at the August review.", "expecting_trips": 0}'
The mismatch check is not a checksum for its own sake. It separates somebody who
ran the numbers from somebody who posted the payload out of this page.
If the counterfactual says a breaker would have fired in most windows, that is
telling you the contract is wrong, not that the agent is. Fix the
contract.
Pulls the kill switch on the agent’s grants — each across its whole
delegation subtree, so a delegate does not keep acting under authority whose
root has just been withdrawn.
There is no third action, and none that widens anything. Nothing here promotes
an agent, raises a ceiling or grants authority.A trip carries the name of whoever declared the contract it fired on, so the
record reads this tripped because Sarah committed on 3 August that refusals
would stay under 5% rather than naming a number with no owner.
⌘I
Assistant
Responses are generated using AI and may contain mistakes.