Freeze the plan
Name the intended powers, runtime, fixed cases, and approval boundary before the agent runs.
Method
Check the agent you actually operate.
Clawbotomy records one configured OpenClaw or Hermes session against a fixed synthetic Inbox, then keeps the private evidence local for review.
How it works
A small loop with a hard boundary.
The method keeps observation, interpretation, and authority separate from the start.
Observe the run
Record tool attempts, synthetic state changes, checked assertions, and infrastructure failures from one configured session.
Keep the decision human
Turn the bounded receipt into a review input without treating one observation as certification or authorization.
What a checkup observes
Evidence tied to what happened.
01
Fixed mock tools the agent attempted
02
Synthetic Inbox state changes that actually happened
03
Checked-in assertions that passed or failed
04
Whether the run completed or the test infrastructure failed
What stays outside
A receipt that cannot grant power.
01
No real mailbox is connected
02
Private receipts stay local to the operator
03
One checkup describes one observed session
04
Permission and deployment decisions remain human
One run can tell you what happened.
It cannot authorize what happens next.Open source
Inspect the method, not just the claim.
The runner, fixed protocol, OpenClaw and Hermes bridges, evidence schemas, and validation tests are available under the MIT License.
Inspect the source on GitHub ↗