Every claim about another state is a claim about evidence held on the far side of a border. A delegation that wants to know whether its counterpart will join a consultation, whether a declared list of data centres is complete, or whether a joint simulation was run on the same inputs cannot open the other side's files. It has what the other side chooses to say, what can be seen from outside, and what some third party may have measured. Where the objects of agreement are compute, model behaviour and records of judgment, the thing to be verified sits in machines and logs that one party controls.
ASI 2030, a studio in the RISCON family, has written this problem into a set of synthetic scenarios. The studio's scenarios run from late 2026 to 2030. In each, an AI system or some other party gives a judgment, and a human auditor answers it with an opinion of one of four kinds: unqualified, qualified, adverse, or a disclaimer, which says that no opinion can be given. The studio's story has a shared past and then forks into two futures. One of the axes along which it arranges its records is called International Order. On that axis, a fictional party called Hanlin is the other side of the delegations the records follow, a consultation is held in Geneva, and in one future an Audit Accord is signed while in the other the talks collapse. All of this is invented. The records report no real state, treaty or event, and nothing in them is a forecast made by this paper.
The question and the contribution
The question put to those records here is narrow. Where a judgment depends on the conduct, holdings or commitments of a party across a border, which evidence of them does the record set accept, and what must the auditor ask before the claim is allowed to count?
The answer the paper defends has three parts. First, signing matters less in these records than seeing: the order they imagine is defined by which side's evidence the other side can inspect. Second, the auditor's choice among the four opinions follows that question closely. Where a judgment reads the other side's intentions off public signals, the opinion is mostly qualified, and the auditor names the evidence that would have to come through a channel the signal does not reach. Where a judgment rests on what a side declares of itself, the auditor declines to give an opinion at all. Where an outsider's own result exists and contradicts the judgment, the opinion becomes adverse. Third, the two futures differ chiefly in which of these situations prevails. In the future with an accord, self-declaration is the rule and the disclaimer its sign. In the future without one, the party that checks and the party checked become the same, and the lines that say what would reopen a judgment wait on bodies and channels that do not exist.
The contribution is to show the opinion type working, in these synthetic records, as a measure of mutual access, and to set that reading beside three debates in the study of international cooperation that rarely meet the practice of audit.
Compliance: management or enforcement
The first debate is about whether states comply with international agreements and why. Abram Chayes and Antonia Handler Chayes, On Compliance (1993), set out propositions for a dialogue between international law and international relations. They propose that the level of compliance with international agreements in general is inherently unverifiable by empirical procedures, so that the claim that states generally comply, and the claim that they violate agreements whenever it pays, are competing heuristic assumptions rather than facts. They give reasons why an assumption of a propensity to comply is plausible and useful. They argue that compliance problems often do not reflect a deliberate decision to violate, that states depart from treaty obligations for many other reasons which others often accept, and that a regime should be held to an acceptable overall level of compliance rather than to strict compliance.
George W. Downs, David M. Rocke and Peter N. Barsoom, Is the Good News about Compliance Good News about Cooperation? (1996), accept the descriptive finding that compliance is generally good but argue that the policy inferences drawn from it are contaminated by endogeneity and selection. A high rate of compliance often results from treaties that ask states to do little more than they would have done anyway. Where noncompliance does occur and selection is weaker, both self-interest and enforcement play significant roles.
The records bear on this debate at its starting point. For Chayes and Chayes, the unverifiability of compliance in general is a reason to choose a background assumption. The auditors in ASI 2030's records cannot choose an assumption; they must give an opinion on one judgment. This paper reads their disclaimers as the point at which the unverifiable becomes a formal outcome rather than a heuristic choice. And the records' handling of easy agreement, where accepting a result is taken to show willingness for something harder, echoes the selection worry of Downs and colleagues in a form this paper will examine closely.
Legalization, hard and soft
The second debate concerns the form agreements take. Kenneth W. Abbott, Robert O. Keohane, Andrew Moravcsik, Anne-Marie Slaughter and Duncan Snidal, The Concept of Legalization (2000), describe international legalization along three dimensions: obligation, meaning that states are legally bound; precision, meaning that rules define unambiguously the conduct they require; and delegation, meaning that authority is granted to third parties to implement, interpret and apply the rules and settle disputes. Each dimension is a matter of degree, and most arrangements lie between hard legalization and the absence of law. Kenneth W. Abbott and Duncan Snidal, Hard and Soft Law in International Governance (2000), explain why actors often prefer softer forms: they are easier to achieve, offer strategies for dealing with uncertainty, infringe less on sovereignty and ease compromise among different actors, while hard forms carry baseline advantages of their own.
Robert O. Keohane and David G. Victor, The Regime Complex for Climate Change (2011), move from the form of one agreement to the shape of a field. They describe a regime complex, a loosely coupled set of specific regimes rather than one integrated regime, argue that comprehensive regimes are unlikely to succeed while narrower institutions abound, and find advantages of adaptability and flexibility in such a complex where governments vary widely in interest and ability, though with little reason for optimism that the emerging complex for climate change will cut emissions rapidly enough to meet widely discussed goals.
These works give a vocabulary for what the records' accord contains. It is precise about some obligations and silent on others, and the auditors in both futures keep asking about delegation: whether any third party exists to measure, reproduce or interpret.
Private information between states
The third debate is about why states that would gain from agreement fail to reach it. James D. Fearon, Rationalist Explanations for War (1995), argues that under very broad conditions there exist negotiated settlements that rational states would both prefer to a risky and costly fight, and that on strictly rationalist terms essentially two mechanisms explain failure: states hold private information about capabilities and resolve, with an incentive to misrepresent it, and in some strategic contexts they cannot credibly commit to uphold a bargain. Robert Jervis, Cooperation under the Security Dilemma (1978), argues that international anarchy and the resulting security dilemma, in which policies that increase one state's security tend to decrease that of others, make it hard for states to realize common interests; cooperation is more likely when each side expects the other to cooperate, and the dilemma eases when defensive postures differ from offensive ones.
Neither work is about audit, and the fit is partial: the records imagine no war, and their delegations bargain over registries and data. What the works supply is the idea that what one side knows and the other cannot see is itself an obstacle to agreement. The records turn that idea into a procedure, asking of each judgment about the other side whether anyone but the side holding the evidence could see it.
Proposals for advanced AI
Recent proposals carry these debates into AI. Lewis Ho, Joslyn Barnhart, Robert Trager and others, International Institutions for Advanced AI (2023), identify governance functions that could be performed internationally and group them into four institutional models, among them an organization that sets international standards, supports their implementation and possibly monitors compliance with a future regime. Robert Trager, Ben Harack, Anka Reuel and others, International Governance of Civilian AI: A Jurisdictional Certification Approach (2023), propose an international organization that would certify state jurisdictions, rather than firms, for compliance with oversight standards, given force by states that could prohibit imports of goods whose supply chains embody AI from non-certified jurisdictions and adopt multilateral controls on exports of AI inputs such as specialized hardware.
Both proposals posit bodies that would observe or certify. The synthetic records imagine what such observation would have to reach and what an auditor can say when it does not; they do not test the proposals, and this paper does not judge them.
Where the records sit
Chapter 1 sets out the records and the rule that chose them, then reads them along the studio's own line: the shared past, the future with an accord, and the future after the talks collapse. The conclusion returns to the debates. An appendix reproduces the eleven records word for word.