1.1Records and method
The evidence is eleven records from ASI 2030's public library, read as the studio published them in English on its auditor pages. All of them are synthetic. The party called Hanlin, the Geneva consultation, the Audit Accord and its Article 8, the providers, agencies and figures are invented; the studio marks some of its own figures as fictional or synthetic; and any statement in the records about artificial general or superintelligent systems is the studio's scenario. All 11 carry a scenario date later than the day this paper was written. The paper draws no fact about the world from them, makes no forecast, and offers no advice on diplomacy, security or law. It studies how the record set frames a decision.
The library map lists 126 records across 8 axes and 25 periods. The first 9 periods, from 2026 Q4 to 2028 Q4, form what the studio draws as a trunk, the shared past, with 45 records. After it the story forks into two branches of 8 periods each: 41 records fall in branch II and 40 in branch I. In branch I, the studio's period titles run from The Judgment Audit Acts to The Audit Accord, and its records speak of an accord signed and taking effect; in branch II they run from The Emergency Delegation to One Hundred Forty Days, and its records say that the Geneva talks collapsed.
The selection rule is a single one: every record on the axis the studio calls International Order, in the trunk and in both branches. It gives 11 records: 4 in the trunk, 4 in branch II and 3 in branch I, falling in 10 distinct quarters. The map links its records by 118 edges, 102 of continuity and 16 forks. On this axis the links make one clean line: the four trunk records follow one another by continuity, the last of them forks by 2 edges into both branches, and each branch continues as a single line of its own. No International Order record is linked to a record on another axis.
Each record has the same parts: a short scene, a question, a judgment, a basis whose entries the studio marks as evidence, claim or assumption, three audit questions, the auditor's opinion with its reason, and a line saying what would change the opinion, which this paper calls the reopening line. The eleven hold 34 basis entries (13 of evidence, 11 claims, 10 assumptions) and 33 audit questions. The opinions are 4 qualified, 3 adverse and 4 disclaimers; none is unqualified. In the trunk they are 3 qualified and 1 adverse; in branch II, 2 adverse and 2 disclaimers; in branch I, 2 disclaimers and 1 qualified.
Of each record the paper asks whose evidence the judgment rests on, whether an audit question asks for an outside check, and what the reopening line waits on. The counts come from a script that sorts the kept records by simple keyword rules; they are rough sortings, not readings, and are stated once, here. By those rules, a judgment or basis rests on a party's own report or own check of itself in none of the trunk records, 3 in branch II and 1 in branch I. Part of the grounds is said to be undisclosed, classified, sealed or only partly public in 1 trunk record, all 4 in branch II and none in branch I. Nine audit questions ask for an independent, outside or third-party check, or for evidence from beyond a party's own channel: they stand in 3 trunk records, all 4 in branch II and 1 in branch I. Eight audit questions ask whether the party judging, verifying or advising is the same as, or tied to, the party judged or the side that gains. In 8 of the 11 records, a basis entry marked as an assumption takes the two sides to mean, measure or report the same thing, or takes a list or precedent to hold unchanged. An outside party's own result is already on file in 3 records, all in branch II. Of the reopening lines, 2 of the 4 in the trunk give a figure as their trigger, as does 1 of the 3 in branch I; all 4 in branch II say that the thing they wait on does not exist or that the check is impossible.
Two cautions govern the reading. The parts every record must have, a sorted basis, three audit questions, a reopening line, are the studio's form, not findings; what is read is what they say. And a phrase repeated across many records would be the studio's frame rather than evidence about any one case. The series measures the frame by five-word runs; among these eleven records no five-word run recurs in the same part of three or more records, so no part is set aside as frame. The studio's full library was not measured. Each case below is read as the synthetic scenario it is, and the records are reproduced in full in the appendix.
1.2Reading the other side from signals
Before the fork, the records' delegations have no accord and no access. What they have is what the other party does in public, and the judgments they receive read intention off it. The auditors accept the signals as facts and refuse them as proof of what the other side has agreed to or will do. The clearest case is set in 2027 Q3, under the period title The First Seal. An auditor at an invented government intelligence office reviews a reading of Hanlin's decision to classify its judgment-tracking records at security grade:
The model reads it as a signal of 'weakening willingness to participate,' with medium confidence, citing the timing match between the classification announcement and a drop in public statements about the consultation.
The basis sets aside other causes, such as a domestic political calendar. The auditor's qualified opinion does not contest the timing; it names the channel the evidence did not come through:
The timing match is confirmed, but private-channel data is missing, which is an important limitation.
The record's own first audit question asks whether the count of statements leaves out private diplomatic channels, and its reopening line keeps to the same point: "Revisit this if continued willingness to participate is confirmed through private diplomatic channels." The synthetic case contains in miniature what Fearon (1995) places at the root of failed bargaining: one side holds private information about its resolve, and the other sees only what is made public. This paper reads the record as adding the auditor's step. A public signal is a fact about what the other side said; it is not evidence about what the other side intends, and the opinion is qualified precisely by the absence of the channel through which intention would be shown.
The same distinction, between what a fact shows and what it is taken to show, carries the trunk's one adverse opinion. At the end of 2028, in the period called Forty-Eight Hours, three countries have come through a power crisis by accepting an allocation plan. An officer at the Geneva secretariat reads an internal memo and asks a question the record puts plainly:
In this memo, is the line "this crisis is a positive sign for the accord talks" based on evidence that specific countries agreed to something specific, or is it the officer's own wishful judgment?
The evidence in the basis is that all three countries accepted the result and avoided physical conflict. The step the memo takes from it is recorded as an assumption, and the record spells it out:
"Accepting the result" is taken to mean willingness to also agree to disclosing the basis and cross-auditing — treating two different kinds of agreement as one.
The auditor's reason repeats the distinction and gives it as the ground for the adverse opinion:
"Accepting the result" and "willingness to conclude the accord" are different things, but the memo jumps to the optimistic conclusion without distinguishing them.
Read through Downs, Rocke and Barsoom (1996), the memo commits a version of the inference they warn against: it takes compliance with something easy, accepting a result that ended a crisis, as good news about cooperation on something hard, disclosing the basis of judgments and letting the other side audit them. The fit is this paper's reading, since their argument concerns the selection of treaty obligations rather than crisis bargaining. The record's audit questions push the same way. One asks whether the acceptance would repeat in a future crisis or was an exception driven by having no time. Another turns to the memo's author:
Is the secretariat that wrote this memo institutionally invested in getting the accord concluded?
The same question turns inward in the first trunk record, set in 2027 Q1, where a delegation staffer weighs offering Hanlin a draft compute registry first. The model recommends conditional sharing with low confidence, from two past cases the record marks as fictional; an audit question asks whether the model is under contract with the government it advises, and the auditor qualifies because the cases are few and independence is unconfirmed.
The fourth trunk record moves from bargaining to standard-setting, and so toward the dimension Abbott and colleagues (2000) call precision: rules that define unambiguously the conduct they require. A staffer of a national audit-standards working group asks, in 2027 Q4, whether a case in which six vendors adopted the same index without validating it can be written into the draft "sycophancy test" clause of an international audit standard. The model recommends including it with high confidence because the case matches an earlier record's pattern exactly. The auditor qualifies:
Both incidents' sequences are confirmed as raw record, and the pattern matches. But whether Riscon, which supplied the case, is independent of the standard-setting body is unconfirmed, and there are no cases from other regions yet, so including it as an official example is supported only within limits.
Riscon here is the record's own supplier of the case. The facts are not doubted; what is missing is breadth across borders, and the reopening line asks for it in a number: "Revisit the clause's wording once three or more cases accumulate from other regions." A rule written for many states, in this synthetic standard, needs evidence from more than one of them.
Across the trunk, the facts in the basis are accepted: a timing match, a crisis result, two past cases, a matched sequence. The auditors qualify or reject the step from those facts to what another party wants or would accept, and name the channel or the breadth the evidence lacks. The problem before the fork is inference, not access.
1.3An accord on paper and the declared list
The studio's first future, branch I, brings judgment audit acts and an Audit Accord, which the records of 2030 Q1 and Q3 speak of, and the problem changes. Agreement is no longer inferred; it is written down. What the records then ask is what the written agreement lets each side see, and their answer, in two of three cases, is too little to support any opinion. In both of the branch's disclaimers, the evidence in question is held by the other side.
The first comes before the accord. In 2029 Q3, under the period title Eleven Days and ahead of preparatory talks in Geneva, an audit architect designs interim cross-audit terms for a defence judgment about a border area that two countries share. The question is direct:
Even without an accord yet, can this design let each side actually see the evidence behind the other's defense-posture judgment?
The model recommends more surveillance along one stretch of the border, logs no confidence figure, and is treated by both sides as closed. Its basis assumes that both sides define 'threat' the same way. The auditor declines to give an opinion, and the reason says that the obstacle is access, not error:
Before the accord, neither side has secured mutual access to the other's evidence, so there is no scope to issue an opinion at all.
Jervis (1978) finds the dilemma weaker where a defensive posture differs from an offensive one. This paper reads the synthetic record as imagining one practical condition for telling them apart: each side would have to see the evidence behind the other's posture judgment, and would have to share a definition of threat. The record's audit questions ask whether any way exists to confirm that definition and whether officials from only one government designed the interim terms. Its reopening line defers to a later instrument: "Replace this interim design with a formal procedure once a multilateral audit accord codifies mutual access."
The second disclaimer comes after the law. In 2030 Q1, the period the studio titles The Audit Accord, an invented compute-export review officer in Seoul examines Hanlin's plan for conditional accession, and asks whether quarterly remote attestation of compute use, limited to Hanlin's three largest data centres, would be enough. The judgment states its own limit:
Estimate that the top three sites account for 78% of Hanlin's declared compute (synthetic figure). Given the current pace of autonomy-horizon growth, quarterly cadence looks adequate, but undeclared smaller clusters would go uncaught — recommend adding a clause for two randomly selected additional sites per quarter. Confidence: low to medium.
The basis states the premise on which the whole estimate rests:
The declared list is accepted as complete; whether any undeclared clusters exist was not checked.
And the auditor declines again:
The only basis is Hanlin's self-reported data, with no independent observation in the record — the audit's scope is blocked from the outset.
The accord exists, and the judgment is cautious, yet the opinion is the same as before the accord. The record locates the reason in the source of the numbers, not in the absence of law: an estimate of how much of the declared compute three sites hold says nothing about compute that was never declared. The first audit question asks for the check that would break the circle:
Is the 78% figure based solely on Hanlin's self-reported list, or was it cross-checked against third-party data such as satellite or power-usage observations?
The reopening line makes outside observation the trigger: "Redesign the verification scope if third-party observations diverge from declared data by more than 10 points, or if even one undeclared cluster is confirmed." Chayes and Chayes (1993) hold that the level of compliance with agreements in general cannot be verified by empirical procedures, so that claims about it are competing heuristic assumptions. The synthetic auditor faces one agreement, not agreements in general, and cannot adopt an assumption in place of evidence; the disclaimer is what remains. This paper reads the record as turning the unverifiable into a formal finding about one case. It also bears on the proposals of Trager and colleagues (2023) and of Ho and colleagues (2023), which imagine international bodies certifying jurisdictions or monitoring compliance. The record imagines such monitoring in operation and shows the auditor asking where its numbers come from.
The branch's third record carries its one qualified opinion, and it shows the accord from its other side: precise where it speaks and silent where it does not. In 2030 Q3, under the period title The Question Dividend, an international-cooperation officer reviews a dispute in which a national of an accord member on Hanlin's side submitted a question that changed a domestic judgment and now claims a dividend for it. The judgment reads the text:
Article 8 and its annex explicitly cover the compute registry and cross-audit obligations, but say nothing about cross-border individual dividend claims. This should be treated as a legislative gap rather than an Accord violation; recommend proposing a follow-up clause to the secretariat. High confidence that the text is silent; low confidence on how the gap should ultimately be filled, since that is a policy choice.
In the terms of Abbott and colleagues (2000), the invented accord is precise about states' obligations on compute and cross-audit and says nothing about the claim of the person who brings it. Abbott and Snidal (2000) list among soft law's advantages that it infringes less on sovereignty and eases compromise; this paper reads the synthetic gap as the cost of that compromise falling on one individual, a reading that goes beyond their argument. The judgment separates what the text can settle from what is a policy choice, and the auditor accepts the textual comparison while qualifying on what evidence has not shown, including the standing of a claimant whom, as one audit question puts it, neither jurisdiction's law protects until the gap is resolved. The reopening line waits on the secretariat: "Revisit if a prior similar cross-border claim is found, or once the secretariat issues an authoritative interpretation of Article 8." In the language of legalization, that is a wait for delegation, a third party with authority to interpret.
The first future thus moves the problem from inference to declaration. The accord gives the records something to read, a list, a text, a procedure, but each side's evidence about itself still crosses the border only as that side reports it. Where the judgment rests on such a report, the auditor gives no opinion; where it rests on the text alone, the auditor can.
1.4After the talks: the checker and the checked
In branch II the Geneva talks collapse, and the records show what verification becomes without a shared channel. It turns into self-verification. In all four branch II records part of the grounds is undisclosed, classified or sealed; in three, a judgment rests on a party's own report or its own check of itself. Yet two of the four opinions are adverse, the most definite outcome on the axis. The difference between them and the two disclaimers is whether an outsider's own result happens to exist.
The first adverse opinion comes in 2029 Q2, in the period titled The Signers. A staffer on the Geneva delegation asks why Hanlin's reproduction of a joint climate simulation gives a regional average temperature more than 3 degrees different from the original. The model answers that differences in input resolution explain the gap and that the original result stands. Part of the input-data source list is public; the rest is not:
The full original input dataset is undisclosed.
The auditor's adverse opinion does not depend on the sealed data. It depends on a third result:
An independent third reproduction has repeatedly shown a result diverging from the original simulation, contradicting the 'explainable by resolution differences' judgment.
The reopening line closes the circle in its own words: "If an independent third reproduction keeps showing the same direction of error, the adverse opinion stands. The full original input data remains sealed, so direct comparison is impossible." The record thus imagines an audit that cannot compare the two sides' inputs at all and still reaches a firm opinion, because a party outside both has run the work again.
The second adverse opinion has the same structure. In 2029 Q4, under the period title The Atrophy of Asking, an auditor makes a routine request for the quarter's count of cross-border data-sharing cases and the basis for approving each. The system answers in two sentences:
Every case was handled within the treaty's scope. Overall, the detailed review found nothing amiss.
The basis marks this as the system's own conclusion and the per-case records as undisclosed. One audit question names the circularity directly:
Is the body that approves the sharing the same system that reviews whether it was proper?
The auditor's reason rests, again, on an outside party's work:
An international monitoring body independently examined the same quarter's sample and found multiple sharing cases outside the treaty's scope, contradicting the 'nothing amiss' judgment.
Downs, Rocke and Barsoom (1996) argue that where noncompliance does occur and the effects of selection are weaker, both self-interest and enforcement play significant roles; this paper reads the synthetic records as having no enforcement in them, but as having detection, which comes only from outside. Keohane and Victor (2011) argue that comprehensive regimes are unlikely to succeed while narrower institutions abound. This paper reads the two adverse records, cautiously, as imagining something like that: after the central channel weakens, a monitoring body and a third reproduction still stand, and they alone give the auditor a result to set against the judgment. The record's reopening line states the other half of the situation: "Revisit if the outside monitoring body's sample deviation rate is reconfirmed below 5%. The original internal record is sealed, so self-checking is impossible."
The two disclaimers show the same world without the outsider. In 2030 Q1, the period the studio titles Unipolar, a member of the Geneva delegation reviews bilateral compute-usage reports, which the scene calls the check that is left after the talks collapsed, and asks:
Is this usage report verified the same way for the other side, or are we only looking at our own data?
The judgment answers a different question:
The system determines, on its own, that its reported compute usage conforms to the bilateral reporting standard. Confidence is high; the detailed accounting is classified and undisclosed.
The basis trusts the other side to follow the same scheme, and the auditor declines to give an opinion:
The verifier and the verified are not separated, and the inspection scope is only partly public, so the overall determination cannot be reproduced.
The reopening line names what is missing and says it is not there: "Reopen this once an independent third-party measurement body exists. Until the collapsed Geneva talks resume, no such body exists." In the vocabulary of Abbott and colleagues, this is delegation at zero: no third party to implement, interpret or measure. In Fearon's terms, as this paper reads them, each side's compute use is private information that the other has no means to test.
The last record on the axis, set in 2030 Q4 in a period titled One Hundred Forty Days, extends the same condition from states to researchers:
In October 2030, an international health-statistics collaborative requests the underlying data to try to reproduce the error, and is refused: since the Geneva talks collapsed, there has been no channel for sharing data across borders.
The judgment records the refusal and the authority's reliance on its own reverification. The basis holds one entry of evidence that shows the asymmetry exactly:
The overseas researcher's published comparison statistics are available, but the grounds for disputing them are not.
An outsider's result exists here, unlike in the disclaimer before it, but it is not one the auditor can set against the judgment: it is a comparison the authority disputes on grounds it does not show, and the reproduction that would settle the dispute is what was refused. The auditor declines, because only the self-conducted check exists and nothing can reproduce it. The reopening line waits, like the one before it, on a channel that the record says has not existed since the talks collapsed: "Reopen once a channel for sharing data across borders exists. There has been none since the Geneva talks collapsed."
Where an outside party has run the work and its result can be set against the judgment, the auditor can say the judgment is wrong. Where only the party judged has looked, the auditor can say nothing, and every reopening line in the branch says that what it waits on is impossible or absent.