The Space Between the Walls: Why Cross-Agent Commerce Needs Two Standards, Not One

A few weeks ago we published The GUARD Convergence: six frontier labs, working separately, had arrived at the same short list of agent controls — Grant, Usage, Audit, Revocation, Data — and every one of those implementations was single-vendor, non-portable, and impossible to verify from outside. Everyone built the controls. Nobody built the layer between the walls.

This piece is about that between-space, and about a mistake it is easy to make once you see the gap: to assume a single open standard closes it. It does not. Governing one agent well and governing what happens between agents are two different problems, and they need two different standards. One is vertical. The other is horizontal. This is the case for both.

Two problems that look like one

Start with the honest version of the gap, because it is narrower and sharper than "we need interoperability."

The vertical problem is governing a single agent. Who granted it authority, what it is allowed to do, what it actually did, how you withdraw that authority, and what data it may touch. This is GUARD, and it is what F-ACT standardises: a common, checkable grammar for the trust that sits on one agent, so that an agent governed by one vendor can be audited by a party belonging to none of them. F-ACT is the answer to "can I trust this actor?"

The horizontal problem is governing a chain of agents. In any real transaction, authority is not held by one actor — it is delegated. A person authorises an assistant, the assistant engages a booking agent, the booking agent calls a payment agent, the payment agent relies on a merchant's agent. Endorsements flow downstream, hop by hop. And the moment authority moves, a new question appears that F-ACT alone cannot answer: when trust is withdrawn at the top, does it actually fall everywhere below?

That second question is not a variation of the first. You can have a perfectly GUARD-compliant agent at every link and still have a broken chain, because each agent can only vouch for itself. Nobody is responsible for the propagation of a withdrawal across the whole delegation path. That is the horizontal gap, and it needs its own standard. We call it T-RUE — Transitive Revocation of Upstream Endorsements.

The failure F-ACT cannot prevent on its own

Here is the concrete failure, stated plainly, because abstractions hide it.

You can have a perfectly compliant agent at every link and still have a broken chain, because each agent can only vouch for itself.

A principal authorises an agent to arrange something on their behalf. That agent delegates part of the job to a second agent, which delegates to a third. Then the principal revokes — the mandate is cancelled, the authority withdrawn at the source. Under a governance model that stops at each individual agent, the top agent knows it has been revoked. But the second and third agents, two hops downstream, may never hear. They are still holding an endorsement that was valid and no longer is. They keep acting. Money keeps moving on authority that no longer exists.

We have a name for that condition: stranded trust — a withdrawn authority that is still being relied upon somewhere downstream because the withdrawal never reached it. It is not a bug in any one agent. Every agent in the chain can be individually compliant and the chain still fails, because no single agent is accountable for the whole path. Stranded trust is a property of the space between the agents, which is exactly the space no vendor's internal controls reach.

This is why one standard is not enough. F-ACT makes each actor legible. It does not, and cannot, guarantee that a revocation at the origin becomes invalidity at every point it ever touched. That guarantee is a different shape of thing, and it is what T-RUE is for.

What T-RUE actually asks for

T-RUE is a standard for the chain, not the actor. Its principle is a single sentence: when trust is withdrawn upstream, it must fall everywhere downstream. It turns that principle into four properties you can check on an endorsement chain, not take on faith.

Traceable, Transparent, Timely, Trustworthy. The tagline adds the point of it all: and still true — a chain is only worth relying on if a withdrawal at the top is guaranteed to reach the bottom.

And, deliberately, T-RUE mirrors F-ACT's conformance ladder so the two read the same way. Where F-ACT has Declared, Enforced, Provable, T-RUE has T1 Declared (the chain, status endpoint, latency target and fail-closed semantics are all answerable from the document itself), T2 Propagating (the envelope describes an active propagation mechanism and a measured latency), and T3 Provable (it points at a tamper-evident propagation log and a third-party verification endpoint you can actually fetch and check). The honesty is built into the tiers: a document can declare that revocation propagates, but only fetched evidence proves it did.

The precedent that already works — described honestly

Stranded trust: a withdrawn authority still being relied upon downstream because the withdrawal never reached it.

There is one system most people already trust that does exactly this, and it is worth naming carefully, without overclaiming kinship.

When you cancel a payment card, it stops working everywhere, more or less at once — not just at the shop where you first used it. The card network propagates the revocation across a web of parties who do not know each other and do not need to. The card scheme is the forcing function: it made "cancelled means cancelled, everywhere" a property of the network, not of any single merchant's goodwill. That transitive, network-wide revocation is one reason a stranger will accept a card from someone they have never met.

T-RUE is trying to make that same property — a withdrawal at the source becoming invalidity across the whole chain — a checkable standard for delegated AI agents. We should be precise about the difference in status, though. The card networks are decades old, universally adopted, and backed by binding scheme rules. T-RUE is a young open standard, published on 5 September 2026, still forming, and adopted by no one at scale yet. The parallel is to the mechanism, not to the maturity. What the analogy shows is only this: transitive revocation is not exotic. We already depend on it every day, in one domain, and the agent economy currently has no equivalent.

And there is one difference that matters more than any of those, because it is the difference people are most likely to miss. A card network is a toll: it takes a cut of every transaction that crosses it. T-RUE is the opposite of that. It is a free, open, royalty-free protocol — published under an irrevocable patent non-assertion pledge and drafted for anyone to implement without permission or payment — and no fee sits on the revocation itself. Nobody pays a toll for a withdrawal to propagate; nobody pays us per hop. We borrow Visa's mechanism, transitive revocation, and explicitly reject its business model. A trust layer that charged rent on every cross-agent transaction would just be a new wall with a tollbooth in it — the very thing this whole argument is against. The rule of the road is free. What can be earned commercially, by us or by anyone, is the independent, tamper-evident proof that a chain conforms — certification and verification sitting beside the standard, never a gate across it.

"Does your chain speak T-RUE?"

Here is why the two-standard split matters commercially, and not just architecturally.

Today, the question you can ask about an agent is "is it governed?" Tomorrow, in a world where agents transact with other agents on your behalf, the question that actually protects you is different: "if I pull authority at the top, are you certain it stops at the bottom?" That is not a question about any one agent. It is a question about the chain — and it is unanswerable unless the chain speaks a common revocation language.

A standard becomes a forcing function at the moment a relying party can refuse the chain that cannot answer. "Does your agent carry F-ACT?" establishes the actor. "Does your chain speak T-RUE?" establishes that the whole path will fail closed together. When the second question becomes normal — a thing procurement teams, regulators and counterparties simply ask — the horizontal standard stops being a nicety and becomes the price of doing cross-agent business. We are not there. That is the work ahead, stated as work, not as a claim that it is done.

The third leg: R-EAL, briefly

One standard governs the actor. A second governs the chain. You need both.

Two standards govern the actors and the links between them. A third governs what actually flows across the chain: content, claims, the artefacts an agent produces or passes on. That is R-EAL — Reader Evaluation of Authenticity & Legitimacy — an open standard for content credibility, built so that a verdict on whether something is real is itself rigorous, resistant to gaming, reproducible, and legible at a glance.

It belongs to the same family for a simple reason. F-ACT answers can I trust this actor? T-RUE answers does trust still hold along the chain? R-EAL answers is what moved across it real? Actor, chain, content — three checkable layers of the same question, each an open, royalty-free standard rather than a product. R-EAL is its own subject and deserves its own treatment; here it is only the reminder that governing agents well is not complete until you can also govern what they hand each other.

Why two, not one — the honest close

The temptation, once you have seen the gap, is to build one grand standard that does everything. We think that is the wrong instinct. Governing an actor and governing a chain are genuinely different jobs, and a standard that tries to be both tends to do neither cleanly. Two focused standards that mirror each other — same tier language, same royalty-free pledge, same insistence on checkable evidence over declared good intentions — compose better than one that sprawls.

None of this is settled. T-RUE is newly published and unadopted; F-ACT is published and royalty-free but a trust layer only governs anything once it is close to everywhere, and neither standard is there. The core standards are covered by a single Australian provisional patent application, filed 2 February 2026 Australian time — 1 February 2026 in Universal time — which is unexamined and confers no granted rights; it is a dated placeholder, nothing more. And the foundation intended to steward these standards independently is being established, not yet registered. We would rather state all of that plainly than let the argument rest on anything it has not earned.

The argument that does stand on its own is the structural one. Everyone is building governed agents. Almost no one is building the standard for what happens between them when authority moves and is later withdrawn. That space — the space between the walls — is not owned by any vendor, cannot be closed by any single control, and is exactly where cross-agent commerce will either become trustworthy or quietly fail. One standard governs the actor. A second governs the chain. You need both.

Read it, or help build it

All three standards are published in full and royalty-free — not products, standards. If the two-standard argument lands, the next step is not to buy anything; it is to read the specifications and pressure-test them. Read F-ACT for the actor, read T-RUE for the chain, read R-EAL for the content — or start with the standard and see why it is free to implement. The controls already exist, in walled gardens. The open, portable, verifiable layer between them is the part that still needs building — and that is an open invitation, not a closed shop.

Reported and argued from published material; no affiliation with, or endorsement by, any named company or network. The card-network parallel describes a mechanism, not an equivalence of maturity or adoption. The provisional patent application referenced is unexamined and confers no granted rights; the foundation intended to steward these standards independently is being established, not yet registered.