[Audit] Re: Individual draft on resource accounting for agent runs: draft-arsentev-agent-run-metrics

[Evgenii Arsentev <[e-mail]>]

Hi Nicholas,

Thanks, this is helpful. I agree with all three points and I'll fold
them into -02.

On 3.6 - parent_run_id and root_run_id only say where a run's
consumption is counted. They don't show that the child was allowed to
act, and -02 will say so (that is a question for audit records, not
for this format). The non-transitive totals stay as they are. A child
that the parent's delegation step points to, but that has no report,
will be treated as missing, not as zero. A collector that sums totals
then has to mark the tree incomplete and list the missing run_ids. One
thing I'd add - if the totals it does have already exceed the root's
subtree_totals, that assertion is wrong whatever the missing child
holds.

On 9.1 - agreed. A signature binds the bytes and who signed them. It
doesn't make the numbers checked. -02 will say that a signed report is
still the reporter's statement, and that reconciliation with the
provider (or an independent measurement) is a separate check. I'm
thinking of an optional member that only points to a detached signed
record (the format wouldn't define the record itself). Does that fit
what you had in mind?

On the example - yes, I think it would help, and I'd like to do it
together. I can write the first version: a complete parent/child pair,
a missing child report, an attempted double count and a namespace
mismatch, each with the expected collector finding. I'll post it here
as plain JSON first. If it holds up, it goes into -02 as an appendix
with test vectors. The signed-record references I'd leave to you (I
don't want to guess your format).

I'd like to have -02 out before the IETF 127 draft deadline.

Evgenii Arsentev

On Sun, Oct 11, 2026 at 6:40 AM Nicholas Templeman <[e-mail]> wrote:
>
> Evgenii, all,
>
> The parent/root identifiers are useful for correlating a delegated run, but ancestry alone does not establish that the child had authority to perform an action.
>
> For accounting, I would preserve the distinction in section 3.6 between a run's own non-transitive totals and a root's subtree totals. A consumer should not add a root's subtree totals to the descendant reports again. A missing child report should remain missing coverage rather than be silently treated as a zero.
>
> Section 9.1 also matters: the report's values remain assertions of the reporter. A detached signed record can bind the asserted bytes and issuer, while provider reconciliation and independent measurement are separate checks. A signature alone cannot turn an asserted resource total into a verified provider bill.
>
> Would a small synthetic two-agent example help test these boundaries? The proposed cases would include a complete parent/child pair, a missing child report, an attempted double count, and a namespace mismatch, with explicit expected findings and detached signed-record references. That would be an interoperability proposal, not a claim that we have implemented or run it.
>
> Best,
>
> Nicholas Templeman
>
> Council of AI / CSOAI
>
> [e-mail]
>
> _______________________________________________
> Audit mailing list -- [e-mail]
> To unsubscribe send an email to [e-mail]



-- 
Evgenii Arsentev, PhD
Chief Executive Officer, digital health
ORCID 0000-0002-9120-7298 · https://arsentev.ai