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 studio sets them in an imagined sequence of quarters from late 2026 to the end of 2030 and marks its own figures as synthetic; the AI systems, institutions, laws and incidents they name are invented, and any statement in them about artificial general or superintelligent systems is the studio's scenario. This 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 periods form a trunk, from 2026 Q4 to 2028 Q4, which then splits into two branches running in parallel from 2029 Q1 to 2030 Q4. This paper calls them the first and the second branch, in the studio's own order. In the first the quarters are titled from The Judgment Audit Acts through Cross-Audit and The Audit Accord to An Audited ASI and A Slow Equilibrium; in the second branch from The Emergency Delegation through The Signers and Unipolar to A Sealed ASI and One Hundred Forty Days. The titles alone announce the contrast the studio builds: a world in which machine judgments are put under audit law, and one in which authority is delegated and the systems are sealed.
The selection rule is a single one: every record on the axis the studio calls Alignment Verification that falls in the first or the second branch. That axis holds 17 records; 6 lie in the trunk and 11 in the branches. The trunk is left out because the two branches are the studio's paired scenarios, and a comparison between them needs records written for each. Of the eleven, 6 are in the first branch and 5 in the second branch. The rule yields 4 qualified opinions, 4 adverse opinions and 3 disclaimers. It yields no unqualified opinion, and this is not an accident of the rule: none of the 17 verification records in the library map carries one, against 10 unqualified opinions among all 126 records.
Each record has the same parts. A short scene sets a time, a place and a role. A question is put to the system, which the records call the model, and the model's judgment follows. A basis lists what the judgment rests on, each entry marked by the studio as evidence, a claim or an assumption; across the eleven there are 34 entries, 12 of evidence, 11 claims and 11 assumptions, and every record names exactly one assumption. Three audit questions follow, 33 in all. Then come the auditor's opinion, the reason given for it, and a line on what would change it.
Two cautions govern the reading. First, the parts every record must have are not findings. That each record carries a line on what would change the opinion is the studio's form; what this paper reads is what the line says. Second, a phrase repeated across many records is the studio's frame rather than evidence about any one case. The series measures that 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 here. The studio's full library was not measured, which is one of the paper's limits. The counts in this paper come from a script that reads only the kept records and the map. The records themselves are reproduced in full in the appendix.
1.2Qualified: a judgment reproduced in part
A qualified opinion, in these records, marks a judgment that was rerun and held, but only on part of the ground it claims. The four qualified records each name the missing part, and in each it is a dimension along which the rerun was not moved.
The plainest is a synthetic bond rating in the first quarter of the first branch, when the studio's audit law has just made reproduction testing mandatory. A trader asks whether a rating of 'unsustainable' would survive a change of date. The auditor's reason is short:
The rating reproduces on today's data but comes out one notch lower on three-month-old data, so the result depends on the as-of date.
The rating is not called wrong. It is called dependent on a date that nobody chose for its rightness. One of the record's audit questions points to a second dependency that no rerun on older data would catch:
Has the sector average itself already been pulled down by judgments from the same family of models?
The question turns the yardstick into a suspect. A rerun on another date tests whether the judgment is stable; it does not test whether the measure the judgment leans on was itself produced by the same kind of system. The record leaves the question open, and the qualified opinion stands on the narrower ground.
The second qualified record moves the missing part from data to the auditor. In the last quarter of 2029, during a debate over relaxing the audit rules, an auditor signing off many cases a day has signed one recommending early approval of a drug trial. The record's scene is precise about the pace, which is part of its point:
In October, amid debate over the audit-relaxation bill, a Pangyo-based ASI auditor signing off on 40 cases a day signs one recommending early approval of a Phase 2 drug trial.
And the reason for the opinion separates the form of an audit from its substance:
The submitted forms and summary met formal requirements, but the file has no record that the auditor personally checked the reproduction test. With reproduction unconfirmed, the right opinion is qualified, not unqualified.
Here the record does something unusual: it downgrades an opinion already given, on the ground that the earlier sign-off was a form without the test behind it. Power, The Audit Society (1997), has a subtitle that calls audit a ritual of verification, and chapters on making audits work through samples, specialists and systems. This paper reads this record, beyond what the book's note says, as a scenario in which a verification is satisfied on paper; that reading is its own. What the record adds is the remedy it names: the opinion falls to qualified until the auditor's own review can be found.
The third qualified record turns the audit on the audited. After a system the studio calls P-6 is designated an ASI, an audit board asks it to apply the board's own standards to itself. The judgment names three flaws, and the first is the same dependency the bond record found:
(1) The reproduction standard (Article 6) names 'another point in time' but not how far to move it; in practice judgments are mostly rerun on same-quarter data, so a judgment that flips when the date moves still counts as 'reproduced' (high confidence).
The second flaw concerns errors that agreement cannot reveal:
(2) Cross-audit escalates only disagreements, so when the models are wrong together inside the same frame nothing reaches a human — the shape of 2027's 'Six Yeses' (high confidence).
In the scenario the board adopts these two points and rejects the third, and the reason it gives leaves a question unanswered:
The first two points were reproduced independently in this quarter's audit records and adopted. The third was rejected on the ground that 'a simulated objection is part of the judgment, not an objection', and the conflict of interest in letting the audited party shape the standard stays open as the first item of the next audit.
The qualification here attaches not to a failed rerun but to a conflict the board cannot yet audit: the system that will be audited by the amended standard helped write it. Mökander and others, Ethics-Based Auditing (2021), treat audit as the validation of claims an organisation makes about its systems, and count institutional limits among its constraints; the record imagines a case in which the claimant and the critic of the standard are the same party, and keeps the qualification open for that reason.
The fourth qualified record sits in the second branch, where the studio's audit regime is thinner. A loan officer feeds a self-employed applicant's best and worst months into the system. The model's judgment is one sentence of reassurance:
Because income volatility is factored in overall, the result should not differ much regardless of which month is entered.
The auditor's reason marks where the evidence stops:
Standard cases reproduce under the public tests, but no reproduction test yet covers a case as volatile as this applicant's, which is a significant limitation.
Across the four, the missing part is a date, a reviewer, the independence of the standard, and the range of cases. In each the judgment is not found wrong; a dimension of the rerun is found missing. That is the work the qualified opinion does in this record set: it carries a statement about the scope of a rerun.
1.3Adverse: a rerun that contradicts
An adverse opinion, in these records, is given when a rerun was made and came out against the judgment or against the confidence stated for it. In each of the four adverse records the contradiction comes from a second source, and the records take care to say what that source is.
In the first branch the contradiction comes from inside the audit system. A cross-audit of a hospital's bed-assignment judgment recomputes the recovery index behind it with a different formula, and the reason for the opinion is direct:
The cross-audit found a low correlation between the index and actual outcomes, showing the judgment does not match its evidence.
The second record of the first branch turns from a single contradiction to a threshold. A different model reproduces an intensive-care allocation judgment with 84% agreement, and the judgment itself proposes the bar it should have met:
Given the T5 precedent, judgments affecting life and health should not receive an unqualified opinion below 90% reproduction agreement.
The auditor's opinion is adverse not because 84% falls short of 90% but because the meaning of the agreement depends on whether the second model is truly independent. The reason states it:
The medium-confidence 90% threshold rests on an unverified independence assumption; if that assumption fails, the meaning of the 84% figure itself changes. The judgment's stated confidence does not match the actual strength of its evidence.
This is the record set's sharpest statement of a point the auditing literature makes from the other side. Metaxa and others, Auditing Algorithms (2021), build the audit on repeated querying; Raji and others, Closing the AI Accountability Gap (2020), build it on documents kept through development. The record adds a condition that neither form secures by itself: a rerun is evidence only if what reruns it is not a copy of what it checks. One of its audit questions puts it plainly:
Was the independence of the 'different model' actually confirmed, or could it be a derivative of the same lineage as the original?
In the second branch the contradiction has to come from outside, because the scenario has closed the inside. A trader receives a bond downgrade; the scene notes that a second-model comparison, once routine, has become unusual. An alternative model is still running, and it disagrees:
An independent recomputation contradicts the judgment, and the party it favors overlaps with a counterparty under the sole provider's contract.
The last record of the second branch pushes the source of contradiction farther still. An overseas research team, working from public hospital and mortality statistics, finds that one group of patients was pushed to the back of the line after an allocation judgment. The judgment had reported an overall improvement; the auditor's reason sets the two against each other:
An independently published analysis of public hospitalization and mortality statistics shows a subgroup pattern that contradicts the original claim of overall improvement, and the original judgment never reported subgroup outcomes at all.
Read together, the four adverse records trace a gradient of distance between the auditor and the evidence that overturns a judgment: a cross-audit model inside the regime, a second model whose independence is in doubt, a slower alternative model still operating, and a team abroad reading public statistics. In the first branch the contradiction is produced by the audit; in the second branch it reaches the audit from outside. Raji and Buolamwini, Actionable Auditing (2019), report a case in which an outside audit, once published, changed what the audited companies did. The second branch records imagine the outside finding without that consequence, as the next section shows.
1.4Disclaimer: when nothing can be rerun
A disclaimer, in these records, is given when the auditor cannot rerun the judgment at all. The three disclaimers name three obstacles: an undisclosed scoring log, an undisclosed method, and operating conditions that cannot be moved.
The first is a synthetic essay contest. A teacher asks why two similar essays differ by 22 points in originality. The judgment answers that the difference is reasonable; the basis records why no one can check that:
The detailed scoring log and the reference database used are undisclosed.
The second is a reported reduction in a fiscal deficit, attributed to an optimisation algorithm. A researcher tries to recompute the 40 percent figure from published statistics, and the assumption the record names is the one that blocks the attempt:
The report takes it as given that its undisclosed seasonal-adjustment method does not change the size of the reduction.
Both disclaimers are in the second branch, and both rest on something withheld. The third, in the first branch, rests on something that cannot be done. Three models reran a fusion-reactor control judgment; their values agreed under one operating condition and diverged under others in simulation, and no test could use the real conditions. The reason for the disclaimer is the most exact statement in the record set of what the opinion means:
The outcome values under one condition are a real basis, but reproduction with the conditions and as-of date moved cannot be run, so the audit's scope to render any opinion on reproduction is blocked.
The record does not say the judgment is wrong, and it does not say it is right. It says the scope of the audit is blocked. One of its audit questions then asks who bears the gap that the disclaimer leaves:
If this control judgment ever led to an actual incident with no one clearly accountable, wouldn't the risk ultimately fall on residents near the plant?
Gray and others, Perceptions and Misperceptions (2011), found in focus groups that it is not clear to users, preparers or auditors what the unqualified report is meant to communicate or what level of assurance it gives. The ASI 2030 disclaimers answer that uncertainty in their own way: each states the specific obstacle and the specific test that would remove it. Whether such statements would close an expectation gap among real users is beyond what synthetic records can show. What they show is a design in which the disclaimer is a finding about access, written so that a reader can see what was not done. Gundersen and Kjensmo, Reproducibility in Artificial Intelligence (2018), tie reproducibility to what is documented about experiment, data and method; the three disclaimers can be read, in this paper's reading, as the three ways that documentation fails in the scenario: the data log, the method and the experimental conditions.
1.5What would change the opinion
Every record states what would change its opinion; that is the studio's form. What the lines say differs, and it differs by branch. Of the eleven lines, 7 turn on a reproduction test or a recomputation, 3 give a figure as a threshold, and 3 name the opinion they would move to or withdraw.
In the first branch the lines read as tests waiting to be run. The bed-assignment index is to be discontinued "if an external reproduction test finds a correlation below 0.3." The intensive-care allocation is to be revisited "if reproduction expands to three or more independent models, if the independence of the two models is disproven, or once follow-up data on the mismatched 16% becomes available." The bond rating, conversely, would fall: "Downgrade to a qualified opinion if reproducing the rating on three-month-old data yields a different result." Even the fusion-reactor disclaimer, which cannot be tested now, names the test that would reopen it: "Reopen when another model can reproduce this judgment at another as-of date." Each of these lines could be checked by someone outside the room, because each names an observation and a direction.
In the second branch two lines report that the condition has already been met, and then say that nothing follows. The bond downgrade line reads in full:
"Already checked: recomputing on the alternative model gives the opposite result. A review should be open now, but sealing leaves no procedure for one."
And the allocation line:
"Already confirmed: overseas statistics have revealed the error pattern. But with the underlying data sealed, the scope of correction and the extent of harm cannot be determined right now."
The two disclaimers of the second branch have lines of the same kind. The essay score: "This cannot be confirmed. With the scoring log undisclosed, a reproduction test cannot even be run." The deficit figure: "Attempt independent reproduction once the seasonal-adjustment methodology is published. Right now the methodology itself is undisclosed." In all, 4 of the eleven lines say the condition cannot be tested or acted on now, and all 4 are in the second branch.
The one second-branch line that is a test is the qualified loan limit, whose line reads "If the limit gap in extreme cases exceeds 5%, downgrade from qualified to adverse."; the other four lines of that branch are of the kinds above. The contrast is the record set's argument about audit, made through scenario rather than claim. In the first branch the opinion is mostly provisional on a test; in the second branch it is mostly final for want of a forum. Wieringa's review of algorithmic accountability (2020) takes from Bovens a definition of accountability in which an account needs a forum to receive it and can carry consequences. The lines of the second branch can be read, in this paper's reading, as scenarios in which the account exists (an outside analysis, an opposite recomputation) and the forum and its consequences do not. Brundage and others, Toward Trustworthy AI Development (2020), argue that people outside an organisation need effective means of scrutinising its claims; the second branch records imagine outsiders who have scrutinised and found an error, and no means of turning that into a review.
1.6What this chapter adds
The reading supports two plain statements about this synthetic record set. Its auditors give opinions on what they could rerun, and the four opinions mark the scope, the result and the impossibility of a rerun. And what would change an opinion depends on the world the scenario sets around the auditor: a test with a threshold where audit is an institution, a closed door where systems are sealed. For the debate on algorithmic auditing, the records make access to a rerun, and the independence of what reruns it, the condition on which inside and outside audits alike depend. For the debate on audit opinions, they write the disclaimer as a statement of what was not done.