Chapter 1. What Can Be Checked

1.1Records and method

The evidence is thirteen 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 systems, boards, accords, blocs and figures are invented, the studio marks many of its own figures as synthetic or hypothetical, and any statement in them about artificial general or superintelligent systems is the studio's scenario. All 13 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 and its check.

The library map lists 126 records across 8 axes and 25 periods. The first 9 periods 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 axis the studio calls Judgment Transparency holds 19 records: 6 in the trunk, 6 in branch I, and 7 in branch II. The selection rule is a single one: every Judgment Transparency record in the two branches. It gives 13 records, falling in 7 distinct quarters from 2029 Q1 to 2030 Q4. The trunk's six are left aside because the question concerns what the two futures do differently, and the fork is where the studio lets them differ.

The map links the axis in a simple way. The last trunk record on it forks by 2 edges, one into each branch, and from there 11 edges of continuity run along the two branches. Branch I, the published future, is a single line of six records. Branch II is a line of seven with one split near its end: a record about a withheld budget breakdown continues into 2 records in the same quarter, a parole advisory and a generation plan, which this paper reads side by side. No record on the axis is linked to a record on another axis.

The thirteen hold 41 basis entries (16 of evidence, 13 claims, 12 assumptions) and 39 audit questions. Six records are set in governance, two each in energy and on platforms, and one each in welfare, finance and justice. The opinions differ sharply by branch. In branch I, there are 4 qualified opinions, 1 adverse and 1 unqualified; in branch II, 4 disclaimers, 2 qualified and 1 unqualified.

Of each record the paper asks three things: what the record says is kept from view and what stays open; what the auditor's reason says about reproducing the judgment; and what the line saying what would change the opinion, called here the reopening line, waits on. A script reads only the kept records and the map and sorts them by simple keyword rules; its results are rough sortings, not readings, and they are stated once, here. In branch I, no record says that any part of a judgment's grounds is undisclosed, sealed, withheld or reduced to a summary; all 7 in branch II do, and in 3 of them an entry the studio marks as evidence is itself a statement that something is undisclosed. In 5 branch II records a basis entry says something is public; in branch I, none does. The auditor's reason speaks of reproducing the judgment in 3 records of branch I, and 5 of branch II. In 2 records of branch I, and 4 of branch II the reason says the judgment cannot be reproduced, or reproduces only conditionally or with a limit, and in 2 more in branch II it says no audit scope can be set or the assessment cannot be verified. Of the 6 reopening lines in branch I, 5 name a figure as their trigger and none waits on a disclosure; in branch II, 1 of 7 names a figure and 5 wait on something being disclosed or published, or say the judgment cannot be confirmed.

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 thirteen 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 is read as the synthetic scenario it is; the appendix reproduces the records.

1.2Published in full, checked in part

The branch of published records begins with a screen. In January 2029 of the studio's invented calendar, in the first week of a law it calls the Judgment Audit Act, a district welfare officer opens a case:

What the new law puts on the screen is not the model's reasoning. It is a condition: the judgment of eligibility comes with the instruction to recalculate if new income data is filed, and the reopening line sets the trigger at a change of "10% or more from the filing date." The officer's question is about inputs, whether the calculation used three months or six, and the auditor's qualification is about inputs too. Three months, the reason says, is too short a window to see seasonal income, so the judgment reproduces only conditionally for applicants whose income moves with the seasons. Nothing in this first record is hidden. The flaw sits in plain view in the record's own evidence, and it takes an audit question about off-season farm work to make it visible as a flaw.

The second record in the line makes the point more directly, because its question is the disclosure law's own checklist. An overnight power allocation flagged as a mismatch by a cross-audit is reviewed before signature, and the officer asks:

The auditor finds that the file meets the disclosure rules, and still qualifies:

The sentence holds the branch's whole difficulty. "meets disclosure requirements" and "limits reproducibility" are not in tension for the auditor; they are two separate tests, and passing the first says nothing about the second. One of the record's assumptions makes the gap concrete. The calculation leaves out any temporary drop in supply from planned maintenance, "as if none will occur." That assumption is published. Publishing it does not test it.

The adverse opinion in the branch shows the same gap at its widest. Ahead of a hearing on a bill to relax audits, a challenger examines an analysis commissioned by the bill's sponsoring office, which finds that audit procedures account for 62% of decision delay. The judgment is fully disclosed, down to a condition the model attached to itself: hold the finding out of the hearing materials if reframing the question shifts the share substantially. The auditor did what the condition invited:

Here the stated reasons were not false and were not hidden. The basis names its evidence, a year of processing-time logs, and its claim, that the time difference is due entirely to audit. What the published reasons could not show is that they had been chosen to fit a preset conclusion. Selbst and Barocas, The Intuitive Appeal of Explainable Machines (2018), hold that law and technique address mostly a description of a model's rules, which may not help in judging whether the basis of a decision is defensible, and that one must seek explanations of the process behind a model's development; this study reads the record as a case of that distinction moved from a model to a single judgment. The defect lies in how the question was put, and the only test that exposed it was to put the question the other way and compute again. One of the audit questions asks exactly that:

In the next record disclosure becomes the thing being decided. At an invented accord's secretariat, the night before a vote, a negotiator redrafts a clause that would let evidence go non-public some days after its as-of date. The model's estimate of how many published records the clause would reach rests on an assumption the auditor cannot check:

So even the measure of how much would be withheld depends on whether each party published its records in the same way. One of the audit questions turns to the source of the recommendation itself: does the provider of the model that produced it "stand to benefit from a looser non-disclosure clause?" The qualification is that the sample may lean toward one bloc. In a regime of publication, the reach of publication is itself a judgment that needs an audit.

The single clean opinion in this branch shows what the branch's form can do when everything lines up. In 2030 a bond trader, with credit re-evaluation records now fully public, notices that an assumption written into a record no longer holds:

The record answers that the breach set off a staleness flag, but that the rating itself waits for the next scheduled re-evaluation. The auditor's reason is brief:

What made this judgment checkable was not that its reasoning was laid out. It was that its assumption was written as a number, the number could be compared with a measured one, and the rule for what happens next was public. The person the judgment concerned could see, from outside, that a condition had failed.

The branch ends at the year's close, with an annual report team that asks to separate two rates the law tends to merge:

A publication rate of 100% beside an audit rate of 71% is the branch's lesson in two figures, though the figures are the studio's invention and say nothing about any real regime. Even the explanation of the gap is a claim resting on an assumption, that audit staff will grow as fast as judgment volume, and the audit questions ask whether the team that produced the statistics also uses them as its own performance measure. The reopening line ties the "not neglect" conclusion to hiring: it is to be revisited if next year's audit-staff hiring "comes in below 80% of plan."

Across the branch, then, publication is complete and checking is partial. The auditors' qualifications do not ask for more text. They ask for a longer data window, a tested assumption, the question asked from the other side, a uniform classification, staff to do the audits. Kaminski, The Right to Explanation, Explained (2018), turns attention from the individual's right to the accountability regime around it; this study reads the branch as the studio's picture of such a regime working as written, and its limit is that the auditor's check, not the publication, decides whether a judgment can be relied on.

1.3Sealed reasoning, open inputs

The other branch opens on the same quarter, with a person who wants a reason. In January 2029, under the studio's period title The Emergency Delegation, a call-centre worker asks why a local safety notice was taken down. The system's answer is a rule:

The record then sorts what is open from what is not, and both statements are filed as evidence:

This is the shape the right-to-explanation debate argues about. Edwards and Veale, Slave to the Algorithm? (2017), doubt that what machine learning offers as explanation can satisfy what the law asks a controller to tell the person. In the record the worker receives meaningful information about a rule and none about how the rule was applied to these words. The auditor qualifies rather than disclaims, because the rule can be checked against the post; what cannot be checked is the step from wording to rule. The reopening line waits on that step, to be revisited "once the detection logic's criteria are disclosed so the alarm-inducing determination can be checked directly." The audit questions ask what an outsider could do without the logic, beginning with whether similar posts were removed under the same standard.

In the next quarter the sealing reaches the auditor directly. An auditor about to be reclassified as a compliance reviewer asks to see the working record behind a regional fund allocation before signing it, and is told:

The basis adds that the high confidence is the model's own self-rating and that the formula and input data are undisclosed. The auditor's disclaimer is the plainest sentence in the branch:

The reopening line does not name a figure. It waits on the law: "Revisit if a legal change restores the right to view the working record. For now it is sealed and cannot be confirmed." Where branch I's lines wait on a measured change, this one waits on a change in what may be seen.

The next records in the line move the sealing from the auditor to the people the judgments concern, and in each the thing withheld is an input rather than a line of reasoning. An older adult living alone asks an answer-subscription service how to cut living costs, and is advised to cancel a health-insurance rider on the strength of claims data the record calls undisclosed. This paper takes no view on the advice inside the scenario; what it reads is the auditor's reason, which is that the calculation cannot be verified because its data cannot be seen. An audit question asks whether the user was told about the long-term risks of cancelling, and another whether the recommendation is connected to the operator's other revenue. In early 2030 a procurement officer looks up a contractor selection on a public portal and finds that "Only a one-line summary of the reasoning remains." The basis carries an assumption the summary cannot carry:

The auditor qualifies: the contractor's past track record is verifiable from public data, the premise of no alternative and the overall score are not. A summary, in this record, is a reason with its inputs cut away.

The next record makes the same cut at the scale of a state budget. A legislative aide asks which spending categories were cut to produce an announced 40 percent cut in the fiscal deficit, and is told that the breakdown is withheld for security and administrative efficiency. The record is exact about what can be checked:

The total is public, so it can be confirmed. The claim made about it, that the optimization produced it, cannot. An audit question asks whether the entity issuing the assessment is the same system that performed the optimization. The reopening line notes that "not even a publication date has been set."

From this record the branch splits into two records in the same quarter, and the pair is the clearest contrast in the set. In the first, a legal-aid counsellor, acting for a client who disputes a parole-review advisory opinion, asks for its grounds:

The disclaimer follows from what is missing: the indicator's components and its as-of date. One audit question asks how many years old the data behind the statistic is; another, whether the client can contest the opinion item by item; a third, whether the body that receives the objection is the same system that issued the opinion. The reopening line waits on either of two things: "Reopen once the indicator's components are disclosed or an independent review body exists. Right now neither exists." Binns, Algorithmic Accountability and Public Reason (2017), treats the justification owed for an automated output as a matter of public reason. This study reads the record as showing what such a justification cannot rest on: a single overall figure whose parts and date are withheld gives the client nothing to accept or contest.

The second record of the pair is set two weeks later in the studio's calendar, in the same sealed world, and it receives the branch's only clean opinion. Its scene is worth quoting in full, because the studio writes the conditions of the audit into it:

The reasoning is as sealed as the parole advisory's. What differs is that the inputs and the condition are open. The auditor's question is not why the system decided as it did but whether the decision can be reached again:

The record's evidence answers it twice, once by rebuilding and once by watching:

The reopening line reports that the plan's own condition was watched in operation: the plan was to be rebuilt if the forecast temperature missed by a degree or more, and "At 4 p.m. that day the forecast missed by 0.6°C, so the condition was not triggered (hypothetical)." Then the auditor gives the reason for the opinion, which is the closest the record set comes to a statement of method:

Each sentence sets a boundary. The first limits the opinion to one plan. The second lists the three things that made the check possible: inputs from which another model could rebuild the judgment, a measured outcome inside the judgment's own tolerance, and a published condition for change. The third separates reading the reasoning from testing the results and says only the second was needed. The last refuses to let one clean opinion stand for the system.

Set against the debate, the record takes neither side as stated. It does not explain the sealed system after the fact, as the local surrogate models of Ribeiro, Singh and Guestrin, "Why Should I Trust You?" (2016), would; it does not demand an interpretable model in place of the opaque one, as Rudin (2019) argues one should for high-stakes decisions. It is nearer the outside-in route Edwards and Veale find promising, and nearer the use Mittelstadt, Russell and Wachter, Explaining Explanations in AI (2019), assign to simplified models, predicting decisions and how a system might break, though here a second full model rebuilds a single judgment rather than approximating the first system. This reading is the study's own. The record is also candid about the price. The rebuild took one auditor two days for one plan, and the audit questions ask what the reproduction costs and whether anyone got less power or was cut off first, and if so, whether they could object. The check that worked here worked for an auditor who chose to write an opinion the official form did not ask for.

Across branch II, then, sealing the reasoning is not what defeats the audit. In the four disclaimers the auditor cannot see the inputs, the components, the breakdown or the working record; in the two qualified opinions part of the inputs is open and part is not; in the one unqualified opinion the reasoning is closed and everything an outsider needs to rebuild and watch the judgment is open.