Action leaves your process
An agent asks another system to mutate something.
OUTCOME VERIFICATION FOR AGENTIC SYSTEMS
A timeout is not a failure. Prove what happened before retrying.
MachineOutcome demonstrates one reliability rule: when an external action has an ambiguous result, read back the real state before success or retry is trusted.
✓ no credentials
✓ no network calls
✓ no real writes
✓ canonical public verifier
THE FAILURE MODE
Acknowledgement disappears
The request times out after the external system may already have changed.
Naive retry duplicates work
Treating silence as failure can create a second write, deploy, payment or message.
Read reality first
Observed post-state decides whether success or retry is safe.
TRY IT WITHOUT INSTALLING ANYTHING
Change the evidence. Watch the safe action change.
This lab calls the same bounded verifier exported by this public repository. It never contacts a provider and never performs a mutation.
READY
Run an ambiguous action.
You will get an outcome status, reason, safe retry decision, evidence trace and the exact verifier input/output.
WHAT THIS PROVES
Small enough to inspect. Strict enough to fail closed.
The public lab demonstrates one invariant. It is not the private MachineOutcome runtime and it does not claim customer adoption, ROI or universal agent reliability.
Bind evidence
Wrong-task and wrong-attempt evidence cannot prove the action.
Bind pre-state
A stale starting state blocks the mutation boundary.
Read after ambiguity
Transport timeout does not decide the real outcome.
Keep UNKNOWN
Missing or ambiguous reality stays unresolved instead of being guessed into success or failure.