- From: Kenne Ives <kennethives@gmail.com>
- Date: Fri, 7 Aug 2026 15:11:40 -0700
- To: Lars Kersten Kroehl <lars@moltrust.ch>
- Cc: paoladimaio10@googlemail.com, W3C AIKR CG <public-aikr@w3.org>, public-agent-identity@w3.org
- Message-ID: <CAEbYQ4qZdC4H_fHuLnQxER6MoR2gzG68F64uMZZv26gkFwSa5A@mail.gmail.com>
Isaac's integrity-vs-availability split is the right cut, and I'd carry it one artifact further. The same property holds for the evidence a trust decision rests on — the verdict about what an agent connects to, not just who the agent is. If that verdict is content-addressed (named by the hash of its canonical preimage) and signed, its integrity is resolution-independent: a verifier confirms it's the right artifact by hash and checks the signature against a published JWKS, trusting no source. Fetch it from a ledger, an operator domain, IPFS, or an email attachment and the check is identical. What stays location-bound is exactly what Patrick flagged: availability and freshness — is it current, has it been revoked. Those resolve by policy, separately from integrity. This is the shape of the tool-safety attestations I work on (CTEF, and draft-etcheverry-action-ref at the IETF): Ed25519/JWS over an RFC 8785 (JCS) canonicalization, content-addressed to the exact tool definition, so drift since vetting fails the hash and the verdict re-derives offline byte-for-byte. It's been cross-validated across independent implementers. In KR terms, I'd model two properties per artifact, per Isaac — integrity-resolution (content-address vs location) and availability-resolution (ledger / domain / registry) — and apply that split to the identity document and the belief-state/verdict artifacts alike, since the resolution question repeats at each. Happy to contribute the tool-safety-verdict side of that crosswalk to kroftrust.md if useful. Best, Kenne On Thu, Aug 6, 2026 at 9:05 AM Lars Kersten Kroehl <lars@moltrust.ch> wrote: > Greetings AI KR CG, > > I'd like to respond to Paola Di Maio's resolution-dependency note. It > asked each group in the cluster to answer five questions in public: the > resolution path a proposal assumes, its latency budget for agent-to-agent > delegation, its behaviour when the resolution layer degrades, the > justification for a ledger where one is used, and which other groups it has > coordinated with. I build a DID/VC verification layer for autonomous > agents, so the questions apply to us. Below is what the system does today. > > One thing up front. In our deployment the resolution question is not the > same across the system. One verification path resolves on-chain, another > recomputes offline. They answer these questions differently, so I answer > per path where it matters. > > 1. Resolution path. The verdict layer recomputes offline. Verification > takes the decoded public inputs and rederives the result. Nothing is > fetched at verification time — no DID, no mandate, no context. The anchor > is a timestamp over a hash, read only to place the hash in time. A second, > older path resolves on-chain against Base on every verification. That path > carries the dependency the note describes, and until it is gone the system > is not chain-independent. > > 2. Latency budget. The offline path does no network resolution. Call-chain > depth adds no resolution cost there. The recursion raised in section 3 of > the note lands on the on-chain path. We have no measured figures under > realistic chain depth to report. > > 3. Degradation. When the anchor layer is unavailable, partitioned, or in > governance dispute, the offline verdict still recomputes, because it never > needs to reach the anchor. What is lost is the timestamp — the ability to > place a result in time, not the ability to verify it. The on-chain path > degrades the ordinary way: no RPC, no verification. > > 4. Ledger justification. The chain supplies a dated, tamper-evident record > that a hash existed at a time. It does not supply the trust judgement. That > is recomputed. An RFC 3161 timestamp or a transparency log supplies the > same dated existence. For the anchor, a chain is one backend among several. > We have recorded this as an accepted direction: the anchor becomes a > backend abstraction, held in the record rather than the identifier, so the > substrate can change without breaking a DID. This is a direction, not > shipped. Today the write path is Base by default. The reason it matters is > jurisdictional. In some jurisdictions an agent, reseller, or developer has > no legal access to a public chain. A verification path that requires chain > access shuts them out. That is a market boundary. > > 5. Coordination. What I can point to with artefacts behind it: a cross-run > of an action-evidence conformance vector set against PSEA > (draft-yossif-psea), with one axis recomputed by a second and third > implementation before it was frozen, and the same vectors checked against a > composition draft's negative classes. Charter-level coordination across > this cluster is thin, mine included. > > On the paper cited in the source pool: its single-chain passage sets a > design requirement for the certifying layer — an arrangement to avoid — and > does not describe that system's current anchoring. The distinction does not > change the note's argument. > > Hope this helps. > > Best, > Lars > > > Am 03.08.2026 um 14:10 schrieb Paola Di Maio <paola.dimaio@gmail.com>: > > Greetings everyone: > > The Trust Layer rests on explicit representation at a critical junction > where it intersects with DID and the blockchain. > > Several community and work groups chartered work on agent identity, agent > trust establishment, and runtime verification of agent claims. Their scopes > overlap. Their vocabularies converge, and published founding texts diverge > on one engineering property that determines whether any of the proposals > will hold under production load, and none of them names it*: resolution > dependency*: where an agent's identity document is fetched from at the > moment a counterparty needs to verify a claim. > > Three arrangements are in play across the current proposals — resolution > from a distributed ledger, resolution from a domain the operator controls, > and resolution from a registry a third party maintains. Each carries a > different answer to who can make an agent disappear, a different latency > profile at the ninety-ninth percentile, and a different failure mode under > network partition or governance dispute. Charter text as published does not > distinguish among them. > > Distributed ledger technology sits in three distinct positions across this > landscape. One proposal places public-chain receipts directly in scope as > audit anchors. Another borrows ledger vocabulary for an append-only record > model and disclaims blockchain and consensus mechanisms explicitly. A third > carries no chain reference in its own text at all. Separately, an > unaffiliated Zurich operator runs a live W3C VC + DID agent trust layer > anchored on an Ethereum layer-two network. A reader confined to charter > documents would conclude that ledger infrastructure plays a marginal role > in agent identity. Deployment evidence says otherwise. > > The concern is not confined to one consortium. Work on agent identity and > agent trust is under way at the IETF, the FIDO Alliance, the OpenID > Foundation, the Decentralized Identity Foundation, the Linux Foundation, > the Ethereum improvement process, NIST, and several security consortia, and > the same variable goes unstated across most of them. Part II surveys that > landscape. One effort, Ethereum's ERC-8004, answers the question explicitly > and answers it as a hybrid, which establishes that the disclosure asked for > here is achievable and already shipped. > > The knowledge-representation problem is: these charters are formal > representations of systems that are still evolving, and the representation > omits the variable that governs the represented system's behaviour. What > gets modelled is the trust relationship. What gets left unmodelled is the > substrate the relationship depends on to hold at runtime. Ontologies built > on the first without the second may crosswalk cleanly and interoperate > poorly. > > Please give feedback on the draft technical not, send corrections s and > edits via github and a ping to this list > https://github.com/w3c-cg/aikr/blob/main/trust/kroftrust.md > > Kindly redistribute to interested parties *ie, other CG and WGs mailing > lists and chairs who may be able to contribute > to address the open questions > > Thank you > Paola Di Maio > > >
Received on Friday, 7 August 2026 22:11:56 UTC