Discussion
The eighteen records answer the paper's question. The auditors in ASI 2030's invented futures do not treat a published formula as the answer a scored person needs. What they accept, where a change in a score is asked about, is a division of the change between the two parties. Where that split is made, in a designer's comparison of two formulas over the same eight pieces of work, the opinion is clean, and it is the only clean opinion among the five records that ask about such a change. Where it cannot be made, because the input that moved is an anonymous colleague's rating, because the formula is sealed, or because every absence counts the same whatever its cause, the opinion is qualified or withheld. The other three clean opinions answer questions in which no change was asked about: a calculation to re-run, a statute to read, an allocation to recompute.
This reading sets the recourse literature's question one step later than the records put it. Ustun, Spangher and Liu ask what a person must change to reverse a decision; Karimi, Schölkopf and Valera ask how the person gets there. Both questions presuppose a scorer that stays still while the person moves. It is a presupposition Barocas, Selbst and Raghavan count among those that feature-highlighting explanations quietly depend on. The records put it first. In the open future the first designer asks not what to change but whether the designer had changed at all, and the record concedes that nobody had looked at whether the formula moved. The sealed future shows the cost of leaving the question unanswered: there, in two records, the auditor's note on reopening admits that no one can say what would bring a score back up. In the terms of Venkatasubramanian and Alfano, recourse underwrites agency extended over time; this study reads the two linked designer records in the open future, in which one person follows one score from an appeal to a clean breakdown two quarters later, as a picture of that agency in a form the studio could audit, and the single designer record in the sealed future, which ends in a disclaimer before a loan renewal, as its absence.
The records also say where separability stops. A correct calculation settles the person's case and leaves the formula's design open; the auditor of the question dividend writes that whether the ratio is well designed "is a separate question from this case," and sends it to a pattern of 30 or more cases rather than to the person who asked. A process that exists does not settle it either; the auditor of the repeat filers writes that "A process existing does not guarantee a fair outcome." Separability is a test of a single answer; questions of design, notice and who set the weights stay beside it.
In the sealed future the records show another route, which is not a split. When the scorer will not show what moved, a union's dispatch logs or a researcher's sample of payouts show that the scorer's account of the people does not match what others recorded about them. Read beside Lee, Kusbit, Metsky and Dabbish, and Calo and Rosenblat, this study takes it as workers' shared sense-making answering an asymmetry of information; the auditors treat it as enough for an adverse opinion, though it tells the person nothing to change.
None of this tells any scorer, worker, court, lender or hospital what to do; the claims concern how a synthetic record set frames a scored person's question.
Contributions
Three contributions follow. The first is the separable score: a test, prior to recourse, of whether a score's movement can be divided between the scored and the scorer; on these two axes the one clean opinion on a change in a score follows it. The second is that publication and separability come apart in these records. An open formula still ends in a disclaimer where some inputs are colleagues' sealed ratings; a formula never published still gets a remedy that strips out the scorer's share; and the sealed future's only clean opinion goes to an allocation published in full and recomputed by someone else, where no change was asked about. The third is that in the sealed future the auditors' lines on what would change their opinion stop giving thresholds and start recording waits, and that two of them state outright that nobody knows what would raise a score; the record set writes the absence of recourse into the place where reopening should be.
Limits
The evidence is eighteen synthetic records from one studio, in the English versions it published, and their auditors' opinions are the studio's own. The sortings in the method section come from keyword rules; each count was read against the records, but another reader could draw the categories differently. The frame was measured only among these eighteen, not across the studio's library. The records give no view of how real scoring systems behave, and the paper draws none. The works named are used as lenses within what their reading notes say, and where the paper goes further, it says so.
Further research
Three lines follow. The four lines of records read here fork from two records of the studio's shared past that an earlier paper in this series read; reading each pair of branch lines back to its common trunk record would show how the same starting scene of a scored person diverges. The test of separability could be applied to records on other axes, where a judgment changes over time and the question is whether the change is the system's or the world's. And the reopening lines invite a measure of their own: whether a line names a threshold a person could watch, or a disclosure the person must wait for, could be counted across the studio's whole library and across other studios that write such lines.
The paper's contribution, restated at the end as at the start, is a single question placed before recourse: me or the formula. In ASI 2030's records, the answer that a scored person can use is the one that tells the two apart.