When something is not there
The behaviour that decides whether an integration is worth having is not what it does when everything works. It is what it does when the source is down, has been changed without notice, or sends something that cannot be believed.
Why each of these is the behaviour and not another
- A failed cycle is still a recorded cycleAn import that silently did not run looks exactly like a source with no new data. Recording the failure is what makes the difference visible.
- Unexpected schema halts everything from that sourceIncluding the records that would still have fitted. A payload whose shape has changed is evidence that the connector's understanding of the source is out of date, and the rows that happen to still parse are the most dangerous ones to trust.
- A record that cannot be trusted is held, never droppedA dropped record is a fact the Office silently does not know. A held one is a question somebody can answer, and it is kept exactly as it arrived so they can see what came.
- Re-sending the same extract changes nothingRecords are matched on the identifier the source itself uses. A re-sent window updates what is there rather than doubling it — which matters, because extracts do get re-sent and double-counted money is the kind of error nobody finds for months.
And if a system never connects at all
The Office loses no statutory capability. A case can be submitted, accepted, assessed against benchmarks, peer reviewed, decided and certified with no connection to anything — because every fact a case needs can be recorded by an officer who says where they read it. What the Office loses is assurance and effort, not capability, and the record says honestly which it had.