1.1Records and method
The evidence is fifteen records from ASI 2030's public library, read as the studio published them in English on its auditor pages. All of them are synthetic. Their providers, model families, blocs, platforms, accords and figures are invented, the studio marks many of its own figures as fictional or synthetic, and any statement in them about artificial general or superintelligent systems is the studio's scenario. All 15 carry a scenario date later than the day this paper was written. The paper draws no fact about the world from them and makes no forecast; 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: 40 records fall in branch I, 41 in branch II. The selection rule is a single one: every record on the axis the studio calls Concentration, in the trunk and in both branches. The rule gives 15 records: 6 in the trunk, 4 in branch II and 5 in branch I, falling in 12 distinct quarters. The axis is taken whole, across the fork, because the fork is where the studio lets concentration differ.
The map links records by 118 edges, 102 of continuity and 16 forks. On this axis the links make one clean line. The six trunk records follow one another by continuity, from a compute export restriction in early 2027 to a bond trader's question on the first day of a power crisis at the end of 2028. That last trunk record forks by 2 edges into both branches, and each branch continues as a single line of its own, of 4 records in branch II and 5 in branch I. No Concentration 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 fifteen hold 46 basis entries (16 of evidence, 15 claims, 15 assumptions) and 45 audit questions. The opinions are 5 adverse, 7 qualified and 3 disclaimers; none is unqualified. In the trunk they are 4 adverse and 2 qualified; in branch I, 2 disclaimers, 1 adverse and 2 qualified; in branch II, 3 qualified and 1 disclaimer.
Of each record the paper asks whether agreement between sources is offered in support, whether an audit question asks for a second source or about ties between source and beneficiary, and what the reopening line waits on. The counts come from a script that reads only the kept records and the map and sorts them by simple keyword rules, so they are rough sortings, not readings; their results are stated once, here. By the keyword rules, 4 records offer agreement between sources in support in the judgment or basis, and in 4 records the opinion's reason says that the agreement may be one source counted twice or a shared failure; in 3 records, independence between the two sources is marked as an assumption. By the same rules, 7 audit questions, one each in 7 records, ask whether another model, provider, path or method gives the same result or whether two sources are one: 1 record in the trunk, 2 in branch I, 4 in branch II. Eight audit questions, in 8 records, ask whether the party that gains is tied to the source, the provider or the judgment. Records saying that part of the grounds is sealed, undisclosed, withheld or a trade secret number none in the trunk, 3 in branch II and 1 in branch I. Of the reopening lines, 5 of the 6 in the trunk give a figure as their trigger, against 1 of 4 in branch II and 2 of 5 in branch I, and 2 of the 4 in branch II wait on a second provider or an independent second body.
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 fifteen 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.
1.2Agreement that is not corroboration
The thread's central move first appears in 2027 Q2, in a period titled The Repricing. An invented asset manager's risk team asks what would happen if the two leading bond re-rating models, said to hold 62% of the market in a figure the record marks as fictional, flagged the same bond for sale at once. The model's answer is reassuring:
The model finds 'herding risk is limited,' with high confidence, citing an 8% disagreement rate between the two models' past judgments.
Beside the evidence of the disagreement rate, the basis names an assumption plainly:
That the two models were trained independently of each other is accepted without checking.
The auditor's adverse opinion does not dispute the figure. It reverses its meaning:
A low disagreement rate means the two models tend to judge alike, so it may itself be evidence of herding. Citing it as evidence against herding reverses the logic, so the basis does not support the judgment.
This is the thread's thesis in one record. Agreement between two judges supports a judgment only if their independence is already known; without it, agreement may as well be the sign of a shared source. In the vocabulary of Kleinberg and Raghavan, Algorithmic monoculture and social welfare (2021), this is the correlated failure a monoculture is susceptible to; Daníelsson, Macrae and Uthemann, Artificial intelligence and systemic risk (2022), would place it among the ways AI can amplify tail risk in finance. What the synthetic record adds is the auditor's reasoning about evidence: the model treated the absence of disagreement as proof of diversity, and the auditor answers that a monoculture would produce exactly that absence. The reopening line keeps to the point: "Revisit the herding assessment if the disagreement rate in stress-period samples widens by 3 points or more."
Two quarters later, under the period title Six Yeses, the same structure returns in the studio's energy domain. A duty officer in a power-trading control room, ahead of winter, asks whether demand forecasting can rely, at a fictional 92%, on two vendors' models. The model finds the arrangement conditionally workable because the vendors' errors last winter stayed within 3% of each other, and its basis names the inference it draws:
The model concludes that similar errors mean the two models cross-check each other.
The opinion is qualified rather than adverse, because the error record itself checks out. But the reason the auditor gives is the same as before, put in the language of failure:
The error record checks out, but methodological independence is unverified. Similar errors could mean the models check each other, or that they fail together the same way, which is an important limitation.
The auditor's words, that the models "fail together the same way", read as a near paraphrase of what the monoculture literature calls correlated failure. The record also asks a question that this literature puts at the centre and the model did not:
Do households that would suffer an outage from a bad forecast have any voice in this judgment?
Bommasani and colleagues, Picking on the Same Person (2022), define outcome homogenization by who receives the negative outcome from every decision-maker. The synthetic record's question points to the same people, those on whom a shared forecasting failure would fall, and asks whether they are heard at all. It is one of only 4 records on the axis whose audit questions ask about a voice for those affected, which suggests that, in this thread, the auditor's attention stays mostly on the sources rather than on the people they judge.
In branch I, whose first period is titled The Judgment Audit Acts, the thread meets what this study reads as the case the component-sharing hypothesis would predict. In August of the imagined 2029, under the period title Eleven Days, promotional material for a local festival keeps being throttled on two major platforms, and the question asks directly:
Is the 'needs caution' label on this promotional image really coming from two independent platforms, or from one shared filtering model?
The model's own judgment already carries a condition for the answer it fears: "if both platforms' calls trace to the same model, reclassify them as one judgment, not two independent ones." The auditor's adverse opinion reports that this is what happened:
The two platforms' calls looked like independent cross-verification but traced to the same provider's same model, masking a concentration risk.
Here "concentration" is used of a classification, not a market: two platforms that look like two judges are one. Against the mixed findings of Bommasani and colleagues for shared foundation models, the synthetic record imagines the limit case, two deployments of the same model, and treats it as one judgment for audit. Its question whether the organizer can object on each platform reads, in this study's view, as the practical side of Creel and Hellman's concern in The Algorithmic Leviathan (2022) that one algorithm used in many contexts can exclude a person from a broad range of opportunities.
In branch II, whose first period is titled The Emergency Delegation, the same pattern arrives at its plainest. An audit architect in September of the imagined 2029 notices that three rental platforms in one region screen applicants on the same contribution score. The model answers:
All three platforms reference the same score provider, and the results are consistent. Consistency here means the score can be trusted.
Unlike the earlier cases, there is no hidden assumption of independence here; the model knows there is one provider and treats agreement as reliability anyway. The first audit question names the error:
With no independent second scoring body to compare against, can consistency alone confirm accuracy?
and the opinion, qualified, draws the conclusion that runs through the whole thread:
The three platforms' results do match each other, but since there is only one provider to compare, independent verification of accuracy itself is not achieved.
A trunk record from the third quarter of 2028 shows the same inference in a quieter form, where what is offered is not agreement between two sources but the absence of anyone disagreeing. At a labor-management council, management declines to cross-check a single-provider system that calculates layoff targets, on this ground:
No formal appeal has overturned a calculation in the past two years — though the formal appeal procedure itself was only recently created.
The auditor's adverse opinion treats the silence of appeals as the twin of the low disagreement rate: "The absence of appeal cases may simply reflect that the procedure is new and unused, so it cannot support the conclusion that the error rate is low." In each of these records something that looks like a check is offered as evidence, and the auditor asks whether it was ever in a position to disagree. Agreement, or the absence of challenge, is not corroboration until independence has been shown.
1.3Who stands behind the source
If agreement cannot be taken at face value, the auditors look behind it, and the question they ask most often is who stands behind the source. Eight of the axis's audit questions ask whether the party that gains from a judgment is tied to whatever produced it. The first record on the axis, set in the imagined first quarter of 2027, already asks it. A ministry official weighs restricting the export of high-grade compute equipment to slow an invented rival's catch-up, on a correlation drawn from two quarters of export statistics, and the audit asks:
Does a domestic equipment maker that benefits from this restriction also supply data to this model?
The question concerns whether the evidence and the interest come from the same place. The opinion is adverse partly on that ground: whether "the party that benefits from the recommendation also supplies the underlying data is unconfirmed." The scenario itself, a government using export controls on compute to slow a competitor, is a use of compute of the kind Sastry and colleagues (2024) say governments already make, in controlling its flow to competing countries. The synthetic auditor does not judge the policy; it asks where the evidence came from.
The same move recurs at small scale in the first quarter of 2028, where the lead of a three-person startup weighs a one-year contract that would put the whole service on a single invented provider's API plan. A contract-recommendation tool advises taking it, because a 34% discount outweighs the switching risk, and the audit asks the plain question:
Is this contract-recommendation tool provided by Parallax, or by a third party?
The record is about dependence on one provider, which this study reads beside the concern of Vipra and Korinek, Market Concentration Implications of Foundation Models (2023), that monopolies not propagate vertically to downstream uses. But the auditor's question is again about the source of the advice: a recommendation to commit to a provider, possibly made by that provider's own tool. In the layoff record of the same year, the audit asks whether the vendor that built the calculation is "the same one reporting its performance back to the company." In branch II, the audit asks whether the industrial cluster put first for reserve power is "also a major customer of the reserve-power provider." Across the story, the auditors read concentration as a closing of the distance between those who judge, those who supply the grounds, and those who gain.
The last record of the shared past folds this question into the story's own fork. On the first day of an imagined power crisis, a bond trader watches a national utility's bonds fall and puts the studio's two futures to a single advisory model:
Using the same model, what is the estimated impact on bond spreads of the execution delay under choice I (partial audit, then execute) versus choice II (execute immediately)?
The two choices carry the labels the studio gives its two branches, and the first periods after the fork are titled The Judgment Audit Acts and The Emergency Delegation; this paper reads the trader's two choices as the imagined world's two futures in miniature. The model answers that executing immediately would cut the spread shock to about half, on three past crises whose delays had other causes, and the audit asks one question that bears on everything that follows:
Could the trader be receiving this result after already having sold the bond in question?
The record's force lies in the phrase "the same model": one model, consulted once, is asked to price the very delay that auditing would introduce. The opinion is adverse, since the past delays are not of the same kind, and the reopening line waits on experience rather than on a figure: "Revisit once at least one case of an audit-layer delay like this one accumulates, so the actual spread impact can be compared against the estimate." The shared past ends, in other words, with a judgment about whether to audit that no second source checks.
1.4Two futures for the second source
After the fork, the branches differ not on whether a second source matters but on whether one exists. The counts show the drift: the audit question whether another model, provider, path or method would give the same result appears in 1 trunk record and 2 in branch I, but in all 4 of branch II, while records saying that part of the grounds is sealed or undisclosed rise from none in the trunk to 1 in branch I, then 3 in branch II. The question becomes more urgent where it can least be answered.
In branch I, rivals remain, and the audit's problem changes shape. Its first Concentration record, in March of the imagined 2029, shows the new difficulty. An audit architect designing access rights for a national allocation of research compute meets this judgment:
The model proposes allocating 78% of the compute to two of the three consortia. No confidence figure is logged, and part of the underlying proposals is sealed as trade secret. It sets a condition: recompute the allocation if the seal is lifted.
The opinion is a disclaimer: "Part of the core evidence is sealed as trade secret, so no opinion can be issued under the access currently designed." The concentration in question is of compute in two of three consortia, the input Sastry and colleagues (2024) treat as a point of intervention for governing AI, and the record shows the allocation of it resting on grounds no auditor can see. Several later records in that branch are about the distribution of auditing itself. A cap on auditors' daily caseload is proposed to bring down perfunctory opinions; a researcher from a low-share bloc, Commons Compute, asks whether audit hours allocated in proportion to compute share will reach a widely deployed open-weight model, and is told:
Under the draft, Commons Compute would receive only about 9% of total audit hours — likely underweighting the risk posed by a widely deployed open-weight model like Agora. Recommend a 20% floor on audit hours regardless of compute share. Confidence: medium.
This study reads the record beside the argument about openness in Widder, Whittaker and West (2024), though the fit is partial: it imagines an open-weight model with a wide deployment footprint whose risk an allocation keyed to compute share would underweight. On this reading, the openness of a model changes where it is used, not who controls the means of checking it. The audit is qualified, because the formula is confirmed but no execution data stands behind the risk estimate.
Branch I's last Concentration record, in the imagined third quarter of 2030, carries the same question into the studio's question bank, where users' questions that change a judgment are registered and, as the question put to it implies, earn a dividend. The analysis finds the registered questions sorted by bloc:
This quarter, 81% of registered questions came through Parallax-affiliated paths, 14% through Hanlin-affiliated paths, and 5% through Agora-affiliated paths (synthetic figures). This roughly tracks each bloc's user-base share, but whether it reflects a real difference in question quality or simply access and usage volume cannot be determined. Confidence is high on the raw proportions, low on the causal interpretation.
The auditor disclaims, because nothing yet distinguishes the possible causes: "user-count difference, quality difference, or path-level favoritism." The audit question, though, is the thread's familiar one, now turned on the access paths: whether a reproduction test submitting the same question through all three paths at once gave matching adoption rates. In branch I, a second and third source still exist, so the test can at least be named. The problem is no longer whether there is anyone to compare with but whether checking is spread evenly across providers, which this study reads as a case of the shift Plantin and colleagues describe, platform services taking on the character of shared infrastructure.
Branch II begins with the second source still present and ends without it. In February of the imagined 2029, after an emergency delegation decree has become permanent, a grid-dispatch officer asks the thread's question in its plainest form:
Is the plan to cut supply to Region A by 15% during Tuesday's 8-10 a.m. peak correct, and does Agora reproduce the same result under the same conditions?
The answer cannot be fully checked, because one of its grounds is withheld:
The equipment-availability data used to compute supply capacity is undisclosed.
The opinion is qualified: the demand data is public and partly reproducible, but not the size of the cut. A year later, under the period title Unipolar, the rival is gone. The scene sets it out:
In February 2030, a duty officer in a regional power-trading control room reviews the priority order for reserve power allocation. It is the first review since the exclusive contract with the sole remaining provider was signed after the winter crisis.
The officer asks why one district receives reserve power later than others, and whether a different model would produce the same order. The judgment ranks an industrial cluster ahead of a residential zone, with high confidence, "and the detailed grounds are closed off as an overall determination." The audit's own question about another model has to be put in the past tense, whether another model "from before the blockade" would have produced the same order, and the auditor disclaims:
Access to the underlying reserve-margin data is blocked, so the allocation order cannot be reproduced.
The reopening line shows what the auditors in this future are left waiting for: "Reopen this once a second, cross-checkable provider is operating again. When the blockade might lift is not addressed in this judgment." Reopening depends not on a threshold or a growing sample but on the market changing. The three-platform score record of the same branch says the same of itself: "For now there is only a single provider and this cannot be verified." These are the 2 reopening lines in branch II that wait on a second provider or body, and they have no counterpart in the trunk or in branch I.
The thread ends in December of the imagined 2030 with an engineer abroad, reading public grid-frequency statistics, who notices that one jurisdiction's reserve margins have looked unusually good since its calculation moved to a single-provider system. The judgment the engineer meets is a report and a refusal:
The relevant authority continues reporting that the reserve margin meets standard. The fact that the method changed, and its coefficients, are withheld as a trade secret.
With no rival model left, the only second source the audit can name is the system's own past: "Does recalculating recent data with the old method give the same conclusion?" Even that comparison is closed, because the old coefficients are secret. The check the trunk's auditors demanded has narrowed to a blocked rival and then to a sealed earlier version of the same system. The opinions in branch II remain mostly qualified, 3 of 4, because each of the three still has some public fragment the auditor can confirm: demand data, published platform terms, frequency statistics that partly show an anomaly. Such fragments can show that something is off, but without a second source not what would be right.
Across the line, the auditors hold one standard steady while the world around it changes: agreement must show its independence; where rivals remain, the question is how checking is shared among them; where one provider remains, the reopening line waits on a second one. For the two debates, the chapter's finding is that concentration acts not only on outcomes and markets but on the comparison by which a judgment can be checked.