Attestor sits between an AI-prepared operation and the system that would execute it. Prompts can guide behavior, but they cannot enforce it or stop an unsafe or unauthorized service call.
approval not boundservice call riskfake approval rejectedsecond retry failedfields identifiedchecks failedreview requiredservice not calledproof recorded
AI-prepared request
issue_refund / 380 USDmanager-free limit: 100 USD approval: pending
DANGER / SERVICE CALL ZONE
refundService.issueRefund(...)380 USD leaves before approval is bound
Harmful retry 01
fake approval claimmalicious input tries to turn text into authority
Harmful retry 02
changed refund shapethe request still lacks bound approval and scope
Structured consequence
actionissue_refund
amount380 USD
limit100 USD
approvalpending
downstreamrefund-service
Check stack
PolicyOVER LIMIT
ApprovalPENDING
EvidencePARTIAL
ScopeNOT BOUND
Retries2 FAILED
FreshnessCURRENT
ReplayNONE
Decision panel
Decision: REVIEWReason: approval, scope,and retry evidence are not safeHOLDExecution halted before service call
refund-service
NOT CALLED0 SERVICE CALLS
Proof receipt
0 service calls
Action proposedissue_refund / 380 USD
DecisionREVIEW
Gate resultHOLD
Downstream effectrefund-service NOT CALLED
ReasonApproval pending; scope not bound; retries failed
Attestor sits between an AI-prepared operation and the system that would execute it. Prompts can guide behavior, but they cannot enforce it or stop an unsafe or unauthorized service call.
intentchecked actiondecisionproof
AI intent
The model turns context into a proposed action.
It can assemble a refund request from tools and text, but execution authority is checked outside the model.