What the timeout actually proves
Consider the assumption that a failed connection means an external action never happened. An integration sends a customer message, then loses the response. The provider may already have accepted it. Sending again can duplicate the message; declaring completion can conceal nonacceptance. The same local symptom permits opposing external realities. A timeout establishes a gap in knowledge, not a verdict about execution.
Let evidence govern recovery
The proposed executor would query the original command or provider receipt where supported. Evidence of acceptance must not silently become a claim of delivery. Keep known failure, known completion and unknown outcome distinct, and expose the intended action, attempts and evidence. Retry with a stable idempotency key only when the provider's contract actually makes that retry safe. Otherwise, preserve uncertainty and reconcile. Reversing a local flag cannot unsend an email; a correction is a separate action requiring business authorization.
A hypothetical expiry trap
Imagine a provider that retains both receipts and idempotency records for only twenty-four hours. An outage lasts thirty hours. This is an invented boundary case, not a research finding. A lookup now returns nothing, and reusing the original key may trigger another send because its protection has expired. The new inference is that missing evidence can reflect retention limits rather than nonexecution. Practical rule: verify that lookup history and retry protection still cover the original attempt. If they do not, a missing record cannot justify a fresh send; seek independent evidence or document an authorized recovery decision.
Test the decision, not the optimism
Simulate failure before acceptance, after acceptance and before local receipt storage. Check that each case produces an honest state and a documented recovery decision, including when provider support is insufficient. Bayou's abstract and introductory material distinguish tentative work from stabilized results. Applying that distinction to external actions is an engineering hypothesis, not an ABOLTUS benchmark. ABOLTUS is in development; these proposals are not established product capabilities. Reconsider the assumption only as far as the evidence supports.