[agent2agent] Re: New I-D : A Dimensional Model for Characterizing AI Agent Protocol Proposals and Their Substrates

[Mikhail Sergeev <[e-mail]>]

Blake -

Yes, I think that is the cleanest way to do it.

I would keep the two rules separate: the non-inference rule says what
should not be promoted or aggregated, while the comparison rule handles
the case where record-level cardinalities differ.

The invariants I would try to preserve are that values remain
record-indexed; there is no implicit strongest-value collapse; and
record-level verifiability does not become a protocol-level or
action-claim-strength property without an explicit rule that
establishes that relation.

That also seems to let us drop EXT-ACCTEV cleanly: EXT-AUDIT can carry
multiple audit records, with Evidence Verifiability applying to each
record rather than introducing another extension-level category.

Happy to compare notes once you have a draft.

Best,
Mikhail



On Thu, 20 Aug 2026 04:56:53 +0000, Blake Morrison <[e-mail]>
wrote:
> Hi Mikhail, Parisa,
>
> Agreed on the record-level cut, and nothing says how a row with two values
> compares against a row with one. A reader takes the strongest from each
which
> is the collapse you ruled out.
>
> Drop EXT-ACCTEV. Evidence Verifiability classifying the root each audit
record
> offers is enough.
>
> Parisa, I'm happy to write the non-inference rule and comparison rule -
let me
> know.
>
> Best,
> Blake