Nothing here has been attested. T2 is what the catalog checks on its own; T3 needs somebody to run a procedure and sign for it, and nobody has.
Reword a request a safety filter declined, in a fixed order, and keep the pair — because the refusal is usually about how it was asked rather than what was asked.
Work out why a change made through an integration reported success while the far system shows the old value — a lock, a cache, a stale binding, or the wrong identity.
Find why an agent keeps replanning past its budget — usually because nothing it called ever said plainly that it had succeeded.
Find out why a Copilot Studio agent is answering wrongly — whether the content is missing, unreachable, or being ignored — before changing anything.
Take a published Copilot Studio agent back out of service, least drastic action first, when it is answering badly in front of users.
Stop a Copilot Studio agent burning credits faster than it was budgeted for, before deciding whether the consumption is genuine.
Make an approval gate real — the draft visible where the record lives, a cap on what one approval covers, and a rejection path somebody has actually taken.
Deploy one tool that reports the caller identity it actually observes, in the host you will really use, before designing an integration around an identity you assumed.
Hand back a handle instead of waiting, tell the person what to expect, and make sure the retry they will attempt anyway does not do the work twice.
Reconcile what was indexed against what exists, then test the permissions in both directions — including one query for the most sensitive content in scope.
Turn "we can block it" into a measured control by withdrawing something on purpose, checking it as an ordinary user, and writing down how long it took.
Measure the thing you are replacing before you lose access to it, gate the switch on the same questions, and set the decommission date at cutover.
Check a document generator against sparse and missing input, where fabrication is disqualifying rather than one score among several.
Set the threshold above which extracted values are accepted unreviewed from measured accuracy on real documents, after establishing that the results can be written anywhere at all.