Chapter 1. The Question Before the Judgment

1.1Records and method

The evidence is twelve 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 systems, boards, platforms and agencies they name are invented, and any statement in them about artificial general or superintelligent systems is the studio's scenario. All 12 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 periods form a trunk, from 2026 Q4 to 2028 Q4, which then splits into two branches that run in parallel from 2029 Q1 to 2030 Q4; this paper calls them the first and the second branch, in the studio's order. The selection rule is a single one: every record on the axis the studio calls Who Asks that falls in the first or the second branch. That axis holds 18 records, 6 in the trunk and 6 in each branch. The trunk is left out because the branches are the studio's paired scenarios, and a comparison needs records written for each.

The map also links records to one another, by 118 edges: 102 of continuity, where a later record carries on from an earlier one, and 16 forks, where a trunk record passes into both branches. On this axis the links make a single line. The six trunk records follow one another by continuity, and the last of them forks into the first record of each branch; within each branch, the six records again form one unbroken line. The trunk record at the fork, which this paper does not read, is titled in the map as a question of whether a government chose from options it had framed itself or only between two that a system's preview offered. The studio, in other words, splits its future on a question about who frames the question.

Of the twelve records, 6 lie in each branch. The opinions divide by branch. In the first, 2 are unqualified, 1 qualified and 3 adverse; in the second, 5 are disclaimers and 1 is qualified. The two unqualified opinions are notable in a library where they are scarce: the whole map holds 10, and 6 of those lie in the first branch.

Each record has the same parts. A short scene sets a time, a place and a role. A question is put to the system, and its judgment follows. A basis lists what the judgment rests on, each entry marked by the studio as evidence, a claim or an assumption: 36 entries across the twelve, 12 of each kind, one of each kind per record. Three audit questions follow, then the auditor's opinion, the reason for it, and a line saying what would change the opinion, which this paper calls the reopening line. The twelve fall in 8 distinct quarters and range across education, welfare, health, energy, labor and platforms, and into the domain the studio calls governance.

For each record the paper asks who puts the question and in what form; whether the record carries the person's own question or one the system shaped; whether the audit asks if a person actually reviewed what was approved; and whether the audit asks if the person had somewhere to object. The counts come from a script that reads only the kept records and the map. Its results, stated once here, are these. In the first branch every judgment states a confidence and 3 set a condition for their own revision; in the second, none does either. In the second branch 3 of the six question lines carry a bracketed mark, recording an original question, a question the system supplied, or the note that no question was asked; 2 show one question rewritten into another; and 1 is a request rather than a question. In 3 of the second branch's records the person learns of the decision from a notice or after the fact. Audit questions on whether a person actually reviewed, signed or questioned something stand in 3 records of the first branch and none of the second; audit questions on whether the system shaped or narrowed the question stand in 4 records of the second branch and none of the first; audit questions about a channel for objecting stand in 1 record of the first branch and 5 of the second.

Two cautions govern the reading. The parts every record must have are not findings: that each record carries three audit questions and a reopening line is the studio's form, and what this paper reads 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 twelve 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. The records are reproduced in full in the appendix, and each case below is read as the synthetic scenario it is.

1.2A signature or a reading

In the first branch a person is always near the judgment, and the audit asks whether that presence did any work. The branch opens, in the studio's chronology, with a new public portal for judgment records and a university student who opens one of them, an approval for widening a road, for the first time. The student does not ask whether the road should be widened. The question is about where the question came from:

The system's judgment answers a different question, the one about routes, and it answers it in the branch's usual form, with a confidence and a condition for its own revision. The audit brings the student's question back. Its first audit question asks for a signature, not a route:

The opinion is unqualified, and the reason given is that the log of who set the agenda is disclosed, with the date of the data, and the judgment reproduces. The reopening line makes the signature the hinge: "Hold the judgment if the responsible official's sign-off cannot be confirmed in the agenda-setting record." In this imagined institution, a recommendation about a road stands only as long as a named person can be shown to have put the question. That is the tracing condition of Santoni de Sio and van den Hoven (2018), applied, in this paper's reading, to the beginning of the chain rather than to its end: the outcome must be traceable to a human, and here the record asks for the human who raised the matter.

A signature, though, can be empty, and the branch's most careful record is about exactly that. A duty officer in a power-trading control room is asked to approve the day's dispatch draft, generated automatically under a standing rule:

The officer's discomfort is precise. The draft is not in doubt; the officer is being asked to approve something whose premises someone else chose. The system's answer confirms the draft and then, in the same breath, reveals where the rule came from:

The basis, meanwhile, says something else:

Read together, the three passages show a chain of approvals in which each link vouches for the one before. Today's officer is asked to trust today's draft because it matches a rule; the rule is trusted because a person reviewed it; and that review is assumed rather than shown. The audit refuses the assumption:

The opinion is adverse, for two reasons the auditor states plainly:

This record reads almost as a case study for the second debate. What Wagner, Liable, but Not in Control? (2019), calls quasi-automation, a person included as a rubber stamp in an otherwise automated decision, is here a rule the system drafted and a person signed, now used to make the next person's approval routine. Elish, Moral Crumple Zones (2019), warns that responsibility may land on the nearest human with limited control; the duty officer, asked to approve under a rule that is not theirs, is that human in waiting, and the question the officer puts is, in this paper's reading, a refusal to become one. What the auditor adds is a procedure. The reopening line asks for evidence of reading, not of signing: "Revisit the entire approval procedure if no record of substantive review (time spent, revision history) by the officer can be found." Time spent and revision history are traces that a signature does not leave. Green (2022) asks that any human role be shown, with evidence, to do what it is meant to; the imagined auditor holds the officer's past approval to that test, and, lacking the evidence, gives an adverse opinion on a draft that matched its rule perfectly.

The same branch shows the failure on the side of the request. A union organizer reviews a judgment that followed a platform's request to improve efficiency, and the organizer's question names what the wording hid:

The first audit question asks whether any person did the questioning that the request should have met:

The opinion is adverse because, the auditor finds, the judgment was made behind wording that hid a fee cut. Here the question that started the judgment was not a person's question but a platform's phrase, and the record treats accepting it unexamined as the fault. Parasuraman and Riley, Humans and Automation (1997), describe misuse as over-reliance that can produce decision biases; this paper reads the record as showing a reliance that sits before the output, in the acceptance of a framing.

The branch also shows what oversight looks like when it works. A nurse on night shift reads a recommendation to move a patient up a queue and asks what kind of thing the reason is:

The opinion is unqualified, on the ground that the file keeps measured values apart from the model's estimates and that a rerun agrees. The record is a synthetic scene and says nothing about any real ward. What it shows is a design choice of the studio's: oversight succeeds here because the record lets the person ask the right question, by separating what was measured from what was projected. Goddard and others, Automation Bias (2012), list among the mitigators of automation bias the provision of information rather than a recommendation; in this paper's reading, the nurse's file is that kind of information, and the nurse's question is what it makes possible.

1.3Asking made into a duty and a measure

In the first branch, asking does not stay an individual act. A high-school debate teacher appears in two of the records, continuing a role the studio began earlier, and through that teacher the branch turns questioning into a school subject with a pass mark. The question the teacher brings is about fairness:

The standard is a striking invention. Students are graded not on asking but on asking successfully, on producing a question that changes a machine's judgment. The system answers with the branch's habitual care about its own reach:

The figures are the studio's synthetic numbers, and they describe nothing outside the scenario. What the passage shows is that once asking is rewarded, the right to ask becomes a resource that can be unequally held, and the audit follows that thread to its end, asking whether students from lower-income families, if they are the ones confined to school accounts, have another way to be assessed. The opinion is qualified: the gap is evidenced, but the sample may lean toward schools that volunteered. The reopening line sets numerical bounds, "Redesign the normalization method if a national resample shows the multiplier below 2x or above 5x."

The teacher's second record, in the last quarter of the branch, asks whether any of this matters:

The answer is modest, and the record keeps it modest:

The auditor nonetheless gives an adverse opinion, because the effect is credited to the curriculum with no control group to compare it against, so that the confidence outruns the basis.

Two things follow for the reading. First, the branch that most values asking also audits asking: the claim that questions make a difference is held to the same standard as any other claim, and fails it. Second, the record turns the human role into something countable, a number of adopted questions, and the count is small and local. Endsley, From Here to Autonomy (2017), describes operators whose awareness falls as systems grow more capable and reliable. The first branch, as this paper reads it, is the studio's picture of an institution trying to keep that awareness by training people to ask, and of the same institution admitting that it cannot yet show the training works.

1.4The question replaced

In the second branch the person is still present, but the question is no longer theirs. The six records show, one after another, the forms this takes. The plainest is a classroom where, as the studio sets the scene, debate practice has turned from asking to receiving recommendations, and a student no longer builds arguments at all:

The question line is not a question. Set beside the first branch's teacher, whose students must produce a question that changes a judgment, this student hands the whole task to the system and receives a ranked answer. The audit asks what was lost by asking only for one side:

The auditor disclaims, because the sources and criteria behind the selection are undisclosed. The studio names the quarter The Atrophy of Asking. Parasuraman and Riley (1997) describe abuse as automation introduced without regard to human performance, so that the operator's role becomes a by-product of the automation; this paper reads the classroom as that idea moved from the control room to the formation of a skill. The student's role in the debate has become what is left over once the system has argued.

A shop owner's record shows the question rewritten. The question line records two questions joined by an arrow:

The refined question is more specific: it asks for the input and for a place to object. But the record does not say who refined it, and the judgment that follows names only a rising risk indicator and sends the owner's contest back to the system itself:

One audit question names the danger in the first exchange: that the system handed back a summary to the short question instead of asking what the owner wanted to know. The reopening line places the remedy outside anyone in the scene: "Recalculate once a third credit-data channel opens up. Right now this platform is the only channel there is." A question about where to contest a decision is answered by the system that made it, and the only change that would reopen the case is the opening of a third credit-data channel.

The counselor's record goes one step further: the system has already asked on the person's behalf. The scene describes the discovery:

The supplied question is vague and agreeable; the counselor's question is dated and pointed, and asks who. The audit asks whether the client may frame an objection of their own or is confined to the question supplied, and then names the deeper risk:

This is the one record in the second branch with a qualified opinion rather than a disclaimer, and the reason is the counselor's question. Because the counselor asked when the home visit took place, the public part of the record could confirm that it was six months old; the cut itself still cannot be reproduced, because the weights are undisclosed. The record thus makes a quiet claim, which this paper takes as its central case: in the second branch an auditor can say something only where a person declined the question offered and asked their own. Lee and See, Trust in Automation (2004), write that trust guides reliance where a full understanding of the automation is impractical. A system that supplies the question goes further than asking to be trusted; in this paper's reading, it shapes what there is to understand before anyone tries.

The remaining three records show the question arriving too late or going nowhere. A budget officer learns of a cut from a notification; the question line itself records that no inquiry took place:

The reopening line admits that even the condition for review has no time attached: "But with no prior consultation procedure at all, there is no way to say when that would even be checked." A welfare officer questions the system directly, but only after the applicant has received a notice, and the system's reply closes the matter before it opens:

The officer asked which threshold applied, income or assets; the answer is all of them together, with a forecast that asking again will not help. And a job applicant, screened out automatically, asks a chatbot why:

The audit asks whether the answer's reassuring tone hides the real reason. In each case the person's question is either never asked or is answered in a way that leaves the person nothing further to ask. The auditor disclaims in all three, and for the applicant the reopening line is a bare admission: "This cannot be confirmed."

1.5What the cases show

Across the two branches the person never disappears, and that is what makes the records useful for the debates on human oversight. The difference lies in what the person does. In the first branch the person asks, and the audit checks whether the asking was real: whether a sign-off was a reading, whether a request was questioned, whether a successful question can be shown to have mattered. Where it cannot, the auditor gives an adverse opinion even on a draft that matched its rule. In the second branch the person is handed a summary, a refined question, a suggested question or a notice, and the audit can reach no opinion at all, except in the one record where a person refused the question offered. The records place human control, in this paper's reading, less in the approval of an answer than in the authorship of the question that called for it.