Chapter 1. Numbers That Stand for People

1.1Records and method

The evidence is ten 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 and marks its own figures as fictional or synthetic; the agents, platforms, vendors, courts and task forces they name are invented, and any statement in them about artificial general or superintelligent systems is the studio's scenario. All 10 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 and this paper calls the shared past; after it the story forks into two branches. The trunk holds 45 records. The selection rule is a single one: every trunk record on the axis the studio calls Measuring Human Worth or on the axis it calls Work and Income. The rule gives 10 records, 5 on each axis, and they fall in 7 distinct quarters. The two axes are taken together because they are where the scenario's numbers are numbers about people: what a person's work is worth, what their score says, what their labour adds. Both axes continue past the fork, 15 records on the first and 13 on the second across the whole library; the branch records are left out so that the paper reads one stretch of the story whole.

The map links records 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 these two axes the links make two clean lines. The five worth records follow one another by continuity, from a migration cost in early 2027 to an estimate of AI's share of a productivity gain in the third quarter of 2028, and the last of them forks into both branches. The five work records do the same, from a staffing cut in late 2026 to a reassignment in 2028, and the last of them also forks. None of the ten links to a record on another axis. The two lines are therefore the studio's own threads, and this paper reads each one in order before reading across them.

Each record has the same parts. A short scene sets a time and a role. A question is put, and a judgment follows; in 6 of the ten the judgment is the model's, and in the other 4 it comes from another named party, an operator's dispatch system, a loan review team, an HR tool or an estimate submitted to a task force. A basis lists what the judgment rests on, each entry marked by the studio as evidence, a claim or an assumption: 29 entries in all, 11 of evidence, 9 of claims and 9 of assumptions, because one record lists only two entries. Three audit questions follow, 30 in all, 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 opinions are 4 qualified, 2 adverse, 3 disclaimers and 1 unqualified.

For each record the paper asks four things. What number is at stake, and what is it allowed to decide? What statement lets the number stand for a person, and how does the studio mark that statement? Does the audit ask whether the number holds when computed again? And does the audit ask whether the person counted can see the number or contest it? The counts come from a script that reads only the kept records and the map. Its results, stated once here, are these. Of the 30 audit questions, 10 ask whether a figure holds again, over another window, by another body or method, under a test or on the missing data, and such a question stands in 8 of the ten records. 7 ask who set the figure and whether that party is independent of it. 7 ask whether the person measured can see, contest or be protected from it, in 6 records. Of the reopening lines, 6 give a figure as their trigger and 4 wait on something another party must do or open. Two evidence entries say that there is no evidence, and two assumptions say of themselves that they are untested or unconfirmed.

Two cautions govern the reading. The parts every record must have are not findings: that each judgment has a basis sorted into evidence, claims and assumptions, three audit questions and a reopening line is the studio's form. What this paper reads is what those parts say, and in particular which statement the studio places under which mark. 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 ten 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.

1.2Two lines from a firm's ledger to a person

Both lines begin inside an organization's own accounts and end with a figure that decides something about people outside it. The work line opens in the studio's first quarter, under the period title Unread Approvals, with an HR team and a proposal from an automated agent. The question is put entirely in the firm's arithmetic:

No person appears in that question except as a position. The judgment answers in kind, a cut of 22% resting on a 71% resolution rate, and the reopening line keeps the same arithmetic: "Revisit the staffing cut if the repeat-contact rate exceeds 15%." This study reads the proposal, in the vocabulary of Kellogg, Valentine and Christin, as an instance of discipline by replacing, and the synthetic record presents it as a matter of rates.

At the end of the same line, in the third quarter of 2028 under the period title The Great Reassignment, the same line ends with one of the people who stayed, asking about the number that kept them:

The words "myself included" mark the distance travelled. The question that began as a count of positions has come back as a person asking which number they are. Between the two, the line passes through a working group designing a cross-audit of contribution scores in hiring, a union organizer auditing a dispatch efficiency index that five platforms feed into rider ratings, and a union asking whether low dispatch ranks fall on one age group or signup cohort. Each step carries a number further from the firm that made it and closer to someone who did not.

The worth line follows the same path on a larger scale. It begins with a single proprietor and an accountant deciding how to book a cost after switching models. The figure is the proprietor's own, but the audit already asks who is doing the classifying:

That question announces the theme of the whole line, the gap between whoever sets a number and whoever bears it. The line then moves through a court ruling that a decision can stand without disclosed grounds, a school refused the data behind a learning contribution index, and a credit subsidiary still using a "contribution score" piloted in spring 2027. It ends where the scenario is largest:

The number here stands not for one person but for the labour of a whole sector, and it is meant to set a tax. This study reads Porter's account of the appeal of numbers in business and government as fitting this scene: faced with a choice among three funding proposals, the task force wants a single figure to anchor one of them.

Reading along the lines also shows a number's logic passing between settings, which is what Mennicken and Espeland ask the field to track. The contribution score is the clearest thread. In mid-2027 the working group borrows its design from a fictional credit-scoring precedent; by 2027 Q3 an attorney asks whether a ruling about sealed grounds will reach contribution-score disputes; and by 2028 a "contribution score" is being used to screen loans. The records do not say that this is one and the same score; this study reads the shared name as one logic carried from hiring into lending:

This study reads that drift through Fourcade and Healy's classification situations. Read so, a score built for one market becomes a position that follows a person into another, and the record treats its use as a matter for audit rather than as a neutral convenience. The opinions change as the lines advance. On the work line the first three are qualified, the fourth is a disclaimer and the last is unqualified; on the worth line the first is qualified and the remaining four are two adverse opinions and two disclaimers. The next two sections ask what separates these outcomes.

1.3The equivalence written down

Every number in these synthetic records reaches a person through a statement that says what the number is. A resolution rate is ability; a gain divides into parts; a sample speaks for a whole; one dispute is the same kind as another. The records consistently write these statements down, and they never file them as evidence. The clearest is the last record of the worth line, whose basis says outright what the 61% figure needs to be true:

This is commensuration in the sense of Espeland and Stevens, Commensuration as a Social Process: different contributions placed on one metric so that one can be subtracted from another. What the record adds is that the studio labels the step an assumption and sets it beside the figure, so that a reader sees both the share and the premise that makes a share thinkable. The audit questions then go after the figure's provenance and reproduction, not after the premise; this study notes that the record leaves the premise standing in plain view without testing it.

The same move appears on the work line, where the decisive equivalence is filed as a claim:

The agent's question had asked whether the criterion was ability or seniority. The record's answer is that a rate was read as ability, and it says who did the reading. In the court record the equivalence is legal rather than statistical, and it is split in two:

Here the commensuration is of cases rather than of people, but its effect falls on people: if the two kinds of dispute are one kind, a person contesting a contribution score loses access to the grounds of the decision. The auditor's adverse opinion turns precisely on this step, giving as its reason that the doctrine was extended "on the assumption alone that the case type matches."

Some equivalences are about samples standing for wholes. On the dispatch index, the judgment rests on logs from three of five platforms, and the basis supplies the step that fills the gap:

The audit's reply is one of the sharpest questions in the ten, because it names a way in which shared use of a number can pass for its verification:

This study reads that question as the record's own version of what Espeland and Sauder call reactivity through commensuration: once everyone measures by one index, the index begins to look confirmed by its spread. The fit is partial. Their account concerns how the measured change their conduct, and these records, each a single quarter's snapshot, show almost nothing of how riders or applicants behave in response to a score. What they show is the institutions around the score treating common adoption as a kind of proof, and an auditor refusing that inference.

The last kind of equivalence is the one Kiviat's work bears on most directly. The dispatch operator says that ranks depend only on location and acceptance rate, and the basis adds the step on which that answer depends:

The phrase "nothing on hand confirms it" is the studio marking its own assumption as unconfirmed. The union's question was whether the bottom of the rankings falls on an age group or a signup cohort, attributes a courier cannot change. This study reads the record, beyond what Kiviat's argument itself covers, as staging the condition under which she finds contention arises: a number that may hold people to account for what is not their doing. The audit asks for a test that would separate those attributes from the two the operator names, and the reopening line sets the threshold for revisiting at "a signup-date gap of 5 percentage points or more in dispatch outcomes." In the scenario the auditor does not argue about desert. It proposes to measure whether the number tracks something the person did.

Across the ten records, then, the basis form does something the debate has mostly described from outside: it separates the figure, filed as evidence when there is any, from the statement that lets the figure stand for a person, filed as a claim or an assumption. That separation is the studio's form and not a finding. The finding is where the equivalences fall. In none of the ten is the statement that a number means a person's ability, a sector's share or a case's type placed under evidence.

1.4Holding twice and being seen

In these invented scenarios the auditor's opinions sort by two tests. The first asks whether the number holds when it is computed again. The second asks whether the person it counts can see and contest it. The one unqualified opinion is the record where both are met, and its basis puts them side by side:

The opinion reason names both and nothing else:

Even this clean opinion is kept open. One audit question asks whether the number was fitted to a decision already made:

and the reopening line ties the opinion to both tests at once: "Revisit the criteria if the list's match rate falls below 80% in next quarter's reproduction, or if appeals overturn a cumulative 10% of cases." A falling match rate would fail the first test; a run of successful appeals would show the second test doing its work against the number. This study reads the appeal channel here as a second measuring instrument rather than a courtesy.

Where the first test fails, the auditor will not let a number become a threshold. The credit record is the plainest case:

The audit asks whether the half-input test that the question itself proposes was ever run, and whether rejected applicants can learn which inputs drove the rejection; the opinion is adverse:

Fourcade and Healy write of the facticity of scores, which makes them organizational devices. This study reads the record as showing the auditor declining to grant a score that standing until it survives a deliberate attempt to break it. The robot-tax figure fails the same test more completely, because there is nothing to reproduce:

The auditor here declines to give an opinion at all, and the reopening line would drop the figure from the rate design "if an independent re-estimate differs by more than 10 percentage points, or if the methodology is never disclosed." The second condition already joins the first test to the second: a number whose method cannot be seen cannot be reproduced either.

Where the second test fails, the person counted cannot reach the number, and the auditor has nothing to give an opinion on. The school record is the starkest:

The first audit question shrinks the request to almost nothing and asks whether even that cannot be confirmed:

This is the only one of the ten records whose basis holds no claim, only a contract clause and an assumption that it admits no exception; the judgment is a refusal, and the opinion a disclaimer. The dispatch record on the work line meets the same wall from the union's side:

and its audit questions ask both what share of the requested data the auditor has seen and whether the operator will allow a test that holds location and acceptance rate constant. Kellogg, Valentine and Christin describe workers resisting algorithmic control individually and collectively. This study reads the union that hires its own auditor to draft questions about a dispatch algorithm as one collective form of that resistance, and the record shows its limit in this scenario: the auditor can frame the test but cannot run it without the operator's data.

The court record shows the second test failing at the level of law rather than data. Its first audit question puts the point in a sentence:

The working group's design, earlier on the same line, fails the second test at the level of coverage. It would audit a sample of companies and leave the rest unchecked for a time, and the first audit question asks what that means for the people outside the sample:

That record still receives a qualified opinion, as do the first records of both lines and the dispatch index. This study reads each of them as meeting one test in part: the resolution rate is confirmed by logs but not rerun on another window; the migration costs check out but leave out stabilization; three platforms' logs are in hand but two are missing; the phased design has a precedent but no protection outside the sample. The qualified opinion in these scenarios marks a number that has passed part of one test and is allowed to stand only for the scope it passed.

The pattern across the ten is therefore not a ladder of accuracy. A number can be confirmed by its own logs and still be qualified; it can carry a significant correlation and still draw an adverse opinion. What moves the opinion is whether the number has been made to hold a second time and whether the person it counts stands where they can see it. On the work line both conditions are met by the end of the shared past, in one record. On the worth line the last record meets neither: the shared past ends with a single unreproduced estimate that the task force wants to use for a tax rate, and the studio's map sends that record, and the auditor's refusal to judge it, into both of the futures that follow.