Chapter 1. Me or the Formula

1.1Records and method

The evidence is eighteen 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 platforms, scores, courts, unions, agencies and figures are invented, the studio marks several of its own figures as synthetic, and anything they say about artificial general or superintelligent systems is the studio's scenario. All 18 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 about a scored person and the check made on it.

The library map lists 126 records across 8 axes and 25 periods. The first 9 periods form what the studio draws as a trunk, a shared past, with 45 records; the story then forks into two branches of 8 periods each, with 40 records in the first branch and 41 in the second branch. Two of the axes concern people who are measured: the studio calls them Measuring Human Worth and Work and Income. The first holds 5 records in the trunk and 5 in each branch; the second holds 5 in the trunk and 4 in each branch. The selection rule is a single one: every record on these two axes in the two branches. It gives 18 records, 9 in each branch, falling in 8 distinct quarters from 2029 Q1 to 2030 Q4. The trunk's ten were read by an earlier paper in this series; the counting script checks the eighteen against the 83 records read by the 7 earlier ASI 2030 papers and finds no record of this study among them.

The map ties the eighteen together simply. Each axis in each branch is a single line of records joined by 14 edges of continuity, so the eighteen form 4 lines, and each line begins with a fork from the trunk. The 4 forks leave from just 2 trunk records, one per axis, so the matching lines in the two branches start from the same scene of the shared past and then diverge.

The studio numbers its branches (I, II) without naming them, and this paper calls them the first and the second branch; its period titles and its records say how they differ. The first branch's periods include The Judgment Audit Acts, Cross-Audit, An Audited ASI and The Question Dividend; the second branch's include The Emergency Delegation, The Signers, The Contribution Score, A Sealed ASI and Quiet Exclusion. The names open future (first branch) and sealed future (second branch) are this study's own labels, not the studio's, and their basis is what the records say: in the first branch, 2 of the 9 records say that some part of a judgment's grounds is undisclosed, not public, withheld, closed off, sealed or cannot be disclosed; in the second branch, 7 of the 9 do.

The eighteen hold 55 basis entries (20 of evidence, 17 claims, 18 assumptions) and 54 audit questions. Eight records are set in labor, three on platforms, two in justice, and one each in welfare, finance, education, culture and health. The opinions differ by branch. The first branch has 3 unqualified opinions, 3 qualified, 2 adverse and 1 disclaimer; the second branch has 1 unqualified, 1 qualified, 3 adverse and 4 disclaimers.

Five records ask about a change in a score, a drop or a rise: three in the first branch and two in the second. Of each record the paper asks three things: who asks, and in whose voice; whether the answer lets a change in the score be told apart as the person's doing or the scorer's; 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 keyword rules; its results are rough sortings, not readings, and are stated once, here.

The question is put in the first person singular in 5 records of the first branch and 2 of the second branch; it asks why in 1 record of the first branch and 5 of the second branch. In 4 records of the second branch, an entry the studio marks as evidence is itself a statement that something is undisclosed, and in those 4 it is the only evidence entry there is; in the first branch, no evidence entry says so. The auditor's reason says that the judgment, a correlation, the calculation or the scoring reproduces in 5 records of the first branch and none of the second branch; it says that something cannot be reproduced or verified, or that there is no basis or way to verify it, in none of the first branch and 5 of the second branch. In 2 records of the first branch the reason sets aside a separate question or matter; in 2 of the second branch it rests on data a union or a researcher collected independently, and in 1 on an independent recalculation.

The reopening lines differ most. In the first branch, 5 of the 9 give a figure in digits as the trigger, and 1 waits on a disclosure. In the second branch, none of the 9 gives a figure, and 8 wait on something undisclosed or sealed being disclosed or published, or on an explanation reaching a person; 3 of them open with the sentence "This cannot be confirmed.", and 2 say that it is not even known what would raise or restore the score.

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 eighteen 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; the appendix reproduces the records.

1.2The open future: splitting the change

In the first branch most questions are put in the first person, and the one clean opinion on a change in a score goes to an answer that splits it. Freelance designers are the clearest case. Three of the branch's records are set around a designer's contribution score; the last says it continues from the second, and the first does not say whether its designer is the same person. In the first, the record of the introduction, set in the studio's June 2029, the designer has downloaded the inputs and the judgment record from a public portal, but not the formula, which the scene says is not public. The model puts the drop down to slower responses. Its basis carries an assumption that gives the record away:

The auditor qualifies the opinion on exactly this point:

What the auditor asks for is not the formula as such. The reopening line names the remedy: "If a formula adjustment within the last two months is confirmed, separate out its share of the score drop and re-notify the worker." The scorer's part is to be taken out and the rest handed back to the designer. Barocas, Selbst and Raghavan list stability of the model over time among the hidden assumptions of explanations that point to features; this study reads the record as that assumption written into a basis and marked, by the record itself, as never checked.

The last of the three answers the same question cleanly. Two quarters after the contribution formula was standardized, in the studio's November 2030, the designer asks whether a three-point annual rise is improvement or formula, and the judgment does the arithmetic in the open:

The opinion is unqualified, and the reason says why in terms of the question rather than the method:

This is the separable score in its plainest form. The same work is run through two formulas, and the difference between the formulas is the scorer's share; the difference between the old average and this year's work under the old formula is the person's. The auditor still asks whether the designer picked the eight pieces or drew them at random, and ties the reopening line to the full set of "~40 deliverables", but the split itself is what earns the clean opinion.

Between these two records stands an appeal, the record the last one continues from, and it shows that a public formula does not settle the matter. In the studio's June 2030 the platform's score runs on a published formula, and the question sets it out:

The arithmetic reproduces exactly, and the drop is traced to three new peer reviews. Then the judgment stops:

The auditor disclaims:

Here the formula is public and still the change cannot be split, because the input that moved is itself someone else's judgment of the designer, and that judgment is closed. The record's basis states the order of priorities plainly: the operating policy "puts peer-evaluator anonymity ahead of the right to appeal". In Barocas, Selbst and Raghavan the tension is between making an explanation useful and keeping the model hidden; in this record what is hidden is not the model but the people inside one of its inputs, which is this study's extension of their point rather than theirs. A later record names the change that followed the appeal as "the peer-review weighting adjustment introduced after the I6-era appeal", so the studio's own story has the weighting adjusted after the appeal.

Two further records show that reproduction is necessary but does not finish the auditor's work. A warehouse worker in the studio's October 2029 asks whether a low reassignment score, which cost the worker transition support, measures skill or absences. The model's basis treats every absence alike:

The correlation reproduces, the auditor writes, "but the reason for each absence is missing from the evidence", and the opinion is qualified. The separable score here is not between person and formula but inside the person's own record: an absence for injury or illness and any other absence are different parts of the person. Alkhatib and Bernstein's point is that an algorithm corrects its criteria only once it has already decided, so unusual circumstances are handled badly; the reopening line in this record does the refining after the fact, re-ranking "once the share of absences confirmed as injury- or illness-related is factored in." The parole record of the same branch is similar. The risk scoring reproduces, but the file does not record that the person saw the basis and contested it, so the opinion is qualified, and the reopening line puts that step first: "Reopen the review to guarantee access and rebuttal first, if that step cannot be confirmed."

The cases that end adverse in the first branch are those where the number belongs to whoever set it more than to the people it concerns. A startup's technology chief, in the studio's February 2029, asks whether a recommended layoff of 30% simply echoes a target already set. The auditor finds that it did:

The count of people to cut turns out to be the requester's figure returned with a confidence attached. The other adverse opinion concerns a dividend paid for questions, an institution the studio invents for its 2030. An auditor who was once a delivery worker finds that the top 1% of recipients took 42% of a quarter's payout and that repeat filers are weighted down. A comment period was held and an appeal channel exists, and the auditor is unmoved:

One audit question asks whether the people the formula classifies as repeated filers and weights down are "actually notified of that classification, or do they just quietly receive a lower payout". The person scored cannot ask about a split that nobody tells them exists.

The clean opinions in the first branch go to answers that compute openly and then mark the limit of what the computation settles. An older person who lives alone, helped by a neighbor, asks why the dividend for a question that raised twelve neighbors' care grades is smaller than the twelve follow-up confirmations together. The calculation follows a public split of 30% to the origin and 70% among confirmations, and the auditor gives an unqualified opinion while setting the formula's design aside:

The reopening line sends that separate question to a pattern rather than a person, redesigning the ratio if originators fall below confirmers "across 30 or more similar cases." And when a union organizer asks whether the union may view the basis applied to the riders with the highest refusal rates, the clean opinion goes to a direct reading of a draft statute:

The opinion answers what the scenario's law allows, not whether anyone has used it; an audit question asks "Has any rider actually exercised the individual-access right this quarter, or does the right exist only on paper?" In the open future, then, the auditors accept a calculation that can be repeated and a change that can be split, and they keep separate from it the questions of design, of notice and of use.

1.3The sealed future: asking on someone's behalf

In the second branch the question changes its grammar. It is put in the first person in only 2 of the 9 records, and in 5 of the 9 it asks why. The people who ask are a union organizer, a citizen, a designer, a teacher, an audit architect, the operator of a training center, a nurse, a delivery worker who turns to an organizer for help, and an independent researcher. In this invented future the question is often put on someone else's behalf: an organizer for riders, a teacher for a student, a nurse for a patient.

A freelance designer appears here too, in the studio's July 2029, with a score that fell 62 points before a loan renewal:

The answer refuses both halves of the question:

The only entry the record marks as evidence is a statement that the formula and its items are undisclosed. The auditor disclaims, and the reopening line, which in the first branch was a trigger, here becomes a statement about the person: "This cannot be confirmed. Since the formula is undisclosed, it is not even known what would restore the score." Ustun, Spangher and Liu warn that without recourse a person is denied as long as the model is deployed and lacks the ability to influence a decision that affects their livelihood. The record does not argue that point; it writes it into the place where the auditor says what would change the opinion. The citizen excluded from a court-affiliated mediation program meets the same sentence in another key: the reopening line reads "This cannot be confirmed. Since the scoring items and weights are undisclosed, it is not even known what would raise the score." Karimi, Schölkopf and Valera hold that such explanations show the destination and not the route. In these two records the person is shown neither, and cannot learn what would change the result.

The sealed future also turns silence into data. An audit architect notices that appeals against an automatic cutoff for the bottom 10% of an influence score, used to allocate arts grants, have fallen, and the system explains the fall as satisfaction:

The auditor's question turns the reading round:

The studio names this period The Atrophy of Asking. Within its scenario, a falling count of appeals is read by the scorer as satisfaction, and the auditor cannot test the reading because the formula and the raw survey answers are undisclosed.

Where the formulas are sealed, the checks that hold come from outside. A union organizer in the studio's March 2029 asks why dispatch for some riders has plummeted and whether it falls on one age group. The answer is flat:

The weighting and logs are undisclosed, and one audit question asks whether an independent review is possible at all "Since the same operator runs the dispatch algorithm and answers this question". The opinion is adverse, and it rests on what the union collected:

Calo and Rosenblat describe the platform firm's vantage between the parties and activity that is hidden from view. This study reads the record as the scenario's answer to that asymmetry: the riders' side builds its own record. Lee, Kusbit, Metsky and Dabbish describe drivers making sense together, in online forums, of the algorithm that manages them; in the studio's sealed future the union's logs play that part and go further, because the auditor accepts them as evidence. The last record on the axis, in the studio's December 2030, does the same with an independent researcher's public sample of payout data:

Both of these adverse opinions test the scorer's account from outside. They do not split a change in a score; they show that what the scorer says about the people scored does not match what others recorded about the same people.

The other adverse opinion in the second branch is a judgment about a single patient. A nurse in an intensive care unit, in the studio's August 2030, logs the third case that quarter of a patient with a severe chronic condition ranked behind another marked as more likely to recover, and asks what values went into the formula. The auditor's objection is to the step from group to person:

Binns and the co-authors found generalisation among the concerns people bring to algorithmic decisions; the record gives that concern to the auditor and makes it the ground of an adverse opinion. The record gives no view on how patients should be ranked, and neither does this paper.

A delivery worker flagged as unalignable, in the studio's September 2030, shows the street-level problem of the open future in a harsher form. The cooperation rating counts every refusal and deviation as deliberate:

The auditor disclaims because the logs "do not distinguish force-majeure causes, so the rating cannot be reproduced", and the reopening line names who would have to listen: "Reopen once the explanation channel reaches an actual person. Right now, the same system that issued the rating also receives the explanation." Alkhatib and Bernstein point to mechanisms for recourse in case of error modeled on human bureaucracies; the record sets the condition more bluntly, as a person at the other end of the channel.

One partly open record is qualified, not cleared. The operator of a publicly funded training center, told that the center's effectiveness score fell and its funding with it, can see the graduates' employment filings, which are public, but not the comparison sample. The judgment cuts funding "with the grounds closed off as an overall determination", and the auditor writes that "without the comparison standard and weighting, the full 18 percent cut cannot be reproduced."

The single clean opinion of the sealed future goes to the one record in it with nothing sealed. A career-guidance teacher asks why a student was placed in a manufacturing internship rather than the IT internship listed first. The formula and the slot data are published, and the auditor's reason is short:

Its reopening line is the only one in the second branch that does not wait: "Because every input is public, it can be reproduced at any time." Even here an audit question asks whether the student has "an actual channel to contest the assignment and request a change". In the sealed future, then, an answer is accepted only where the future is not sealed, and elsewhere the auditors record either a contradiction found from outside or the absence of anything that would let the person know what to change.