Blockchain Arbitration & Commerce Society (BACS) · 5 October 2026
A network describing itself as “the adjudication layer for the agentic economy” has launched something called Internet Court. This is neither a metaphor nor a laboratory exercise: the consortium behind it was announced last July with OKX, MetaMask/ConsenSys and Matter Labs among its founding firms, and mainnet launch is expected this quarter.
It deserves attention, because it poses well a question the sector had been asking badly.
What has been built
The mechanism has a name of its own: Optimistic Democracy. When a dispute arises, one validator proposes the answer and the others verify it. What is distinctive is who those validators are: each runs a different, undisclosed language model, and the decision is reached by majority. The stated foundation is Condorcet’s jury theorem, on which a group of independent and reasonably competent decision-makers is right more often than any one of them alone.
A dissatisfied party may appeal within a thirty-minute window by staking gas. The panel then doubles under a 2n+1 rule: five validators, eleven, twenty-three, and so on up to the full set available on the network. The opacity of the models is deliberate — they call it greyboxing — and rests on sound technical reasoning: a heterogeneous panel whose composition nobody knows resists manipulation far better than a single public model.
The contracts governing the system are not ordinary smart contracts. They are programs able to read the web and interpret natural language, so that they can assess claims no deterministic contract could resolve: whether the specification delivered matched what was agreed, whether the flight was “significantly” delayed, whether the event actually occurred.
This is serious engineering answering a real need. For a fifty-cent dispute between two autonomous agents, no court anywhere is an option, and no deterministic contract knows what to do.
The right question
The usual reaction is to ask whether a machine can judge well. It is an interesting question and probably the wrong one: in many disputes of this kind, a panel of diverse models will be as right as a reasonable human being, and will be so in thirty minutes for half a dollar.
The question that decides cases is a different one: is what that panel produces an award?
This is not an academic distinction. The answer determines who holds anything at all once the procedure ends.
Measured against what an award requires
The exercise is worth running against the elements a legal system demands before a private decision binds.
The seat. Every arbitration has a place, and that place does three things: it determines the law governing the procedure, it identifies the court competent to set the award aside, and it gives the award a nationality. The 1958 New York Convention is built on that last function: its Article I speaks of arbitral awards “made in the territory of a State.” A network has no territory. Without a seat there is no lex arbitri, no court of annulment, and no award of any nationality for a second State to recognise.
The identifiable arbitrator. Spanish arbitration law imposes duties of independence and impartiality on the arbitrator and provides a challenge procedure precisely because parties are entitled to know who is judging them and to object. Greyboxing makes that impossible by design: the panel is drawn by lot and the models are not disclosed. This is not an implementation defect a later version might correct; the opacity is the security measure. A decision-maker who cannot be known cannot be challenged.
Form and reasons. An award is made in writing, signed and reasoned. Article IV of the New York Convention requires the party seeking enforcement to produce the duly authenticated original award, and Article V allows the judge to review whether the tribunal was properly composed and whether the party was able to present its case. A tally of votes is not a statement of reasons, and there is no signed instrument to produce or authenticate.
The hearing. Equality, the right to be heard and adversarial process are principles Spanish law declares non-derogable and which the Convention records as a ground for refusing recognition. In the system under examination, validators read transaction data, receipts and signed evidence. Nobody pleads. Nobody is heard.
The symmetry nobody points out
A two-directional conclusion follows, and the second direction matters more than the first.
The losing party has nothing to challenge. There is no award to annul, no court of the seat to approach, no statutory ground to invoke. But the winning party has nothing to enforce either: no instrument to place before a judge, nothing to authenticate, no decision to recognise. The verdict does not exist for the legal order, and is therefore both unassailable and unenforceable.
This is not an objection to the system. It is a description of what it is.
What is genuinely interesting
Because it works regardless, and it works for a reason worth understanding: it does not need the State.
The verdict executes through a controller on the infrastructure itself, which can continue, constrain or revoke the agent’s authority. If an agent acted outside its mandate, its next payment attempt is rejected before the money moves. The decision is self-executing within the perimeter the parties granted it, and inside that perimeter it has no need of a judge.
What we have before us, then, is not defective arbitration. It is a different category, and it is precisely the mirror image of the one we have been describing in these pages.
For a year we have maintained that an award over crypto-assets operates in personam: it obliges a person to do something, but it does not itself move an on-chain asset, nor reverse a confirmed transaction, nor reach someone who controls a set of keys and chooses not to comply. Here the opposite holds. The decision does act on-chain — it is the only thing it knows how to do — but it does not reach a court.
Each has the other’s defect.
Where the limit lies
The system’s own documentation acknowledges this with a candour we appreciate, and it is the most valuable datum in the whole set: the verdict does not reverse a payment already settled, unless that specific transaction was also routed through a separate escrow or refund mechanism.
The perimeter is thus precisely drawn. The agent’s future authority, yes. A transfer already consummated, no.
And the moment value has left, the dispute once again needs a forum with a State behind it. Which is why the expression they themselves use — they describe the network as a first-instance layer, designed to complement rather than replace existing legal frameworks — is more than a courtesy. Whoever says first instance is conceding that a second is missing.
The architecture this implies
None of the above points to competition. It points to a division.
An operational layer resolves quickly, cheaply and self-executingly within the perimeter the parties have handed it. An arbitral layer exists for what crosses that perimeter: amounts that justify the cost, solvent counterparties, assets that have already moved, decisions that must be capable of being placed before a judge. With a seat, with a named and challengeable arbitrator, with a reasoned and signed award, and with the New York Convention behind it.
What is missing is neither of the two layers. What is missing is the engineering of the joint between them: the clause that articulates them, the escrow that lets the first actually bite, the controller that executes it, and the arbitration agreement that opens the second without requiring the first to be litigated again.
That is not protocol work. It is contract drafting, and nobody is doing it today.
Why this exceeds any one project
Machine-produced decisions are going to multiply, and most of them will be reasonable. Their quality will not be the issue. Whether they produce anything a State will enforce, will be.
This is the legal-oracle problem we have set out before in these pages: translating a decision into an executable instruction requires someone who answers for the translation. And it is the problem of Internet Jurisdiction: a functional legal space needs, if it is to be more than a technical one, an anchoring point in a legal order that recognises it.
BACS has no relationship with the project examined here, and nothing above is intended to disparage it: what has been built addresses a problem that existed, and addresses it thoughtfully. Our observation is more modest and more concrete.
They have erected the first instance. The second is not improvised, and is better designed before it is needed.
BACS is a non-profit arbitration institution registered in the EU Transparency Register (REG 9106897105368-14) and operates a Court of Arbitration specialised in digital assets. This article is for informational purposes and does not constitute legal advice.
This analysis is based exclusively on the examined project’s public documentation as at 1 October 2026. BACS has no contractual or other relationship with that project or with the entities forming its consortium.
