[Audit] Re: Records for declined actions [was: Re: Question on AUDIT charter's use case scope]

[Sumit Ahuja <[e-mail]>]

Mirja,

Taking your question directly.

The PR text is a real improvement and "invoked or declined actions" covers the case Henri and Parisa were separating. On whether it needs to be explicit;

Inside a domain, a missing record is a prompt to investigate. An operator can ask the system, read a second log, or call someone. That is the case where leaving it to the group is fine, because the people doing the investigating will work it out.

Outside the domain it is different. The party an inter-domain audit exists for holds the records and cannot reach anyone. For that party a missing record is permanently ambiguous between no action or no record, and no amount of correlation across the records present resolves it, because correlation cannot reach a record that wasn’t there.

That is why the placement matters. If the data models are specified without the boundary stated, they will be built for correlation, and record-worthiness will follow from having a downstream action to correlate against. A declined action has none. It then falls out of scope by construction, which is Henri's point, and by the time that is noticed the models exist.

So the narrow version of the answer: the charter does not need to say what to record. It needs to say that the data models declare their coverage boundary, so a reader can tell "not observed" from "nothing happened". Your new paragraph gets most of the way there. Whether it is good enough depends on whether the boundary obligation is read into "the scope of such auditing is the observable behavior of agents" or has to be said.

I measured this exact failure, which is why I keep raising it. In a pre-registered evaluation of long-horizon agentic serving, two schedulers admitted 107 compliant flows but completed none. No errors, no timeouts, no refusals.

Sumit P. Ahuja
Main Labs