Why two at all
A usage limit on one provider stops work partway through a task. Running a second one keeps a phase moving when the first is unavailable. The reason it is safe to switch is that both are held to the same governance: the same approval gates, the same review model, the same refusal to combine private data with untrusted content and a way out. Which agent did the work does not change how carefully it shipped.
The audit reads, it does not fix
After any change to the shared configuration, an audit compares each file against its own recorded baseline and reports what moved. It never edits either side and never makes the two identical on its own. It describes each difference and proposes the matching change for a person to approve or reject. Automatic synchronization would quietly erase a difference that was there on purpose, which is the opposite of what an audit is for.
The comparison it does not make
The audit checks each file against its own baseline, not one provider against the other. So a rule that is present on one side and simply absent on the other produces no finding: from each file's own point of view, nothing changed. That gap is real, and it is closed by hand, in a single written list of known differences that a person maintains, rather than pretended away. It is the same move as the evidence boundary on the release guides: name exactly where the automation stops and human judgment starts, instead of implying the tool covers more than it does.
Values come from the tool, not from memory
Any recorded fingerprint or reviewed commit is taken from the audit's own output, never typed in from recollection. A value remembered slightly wrong looks identical to one that genuinely checked out, and only one of those two means anything. So the person approves what the tool reports, and does not hand-copy a number it did not produce.