In ASI 2030's two futures, the auditor's check rests in the end on a short list: what went into a judgment, when, under what condition it lapses, and whether someone else can arrive at it. That list, which this paper proposes to call the checkable judgment, is its contribution to the debate over explanation. The records are synthetic scenarios, and nothing here is a claim about any real system, law or regime.
Discussion
The first finding concerns the future of published records. There, disclosure is complete and the check is not. A file can hold its basis, its assumptions, its condition for change, its as-of date and its dissent, as the overnight allocation's file is found to do, and the auditor can still find that it "meets disclosure requirements" yet will not reproduce across a change of season. The adverse opinion in that branch goes further: a judgment whose evidence, claim and condition were all published had echoed the frame of the office that commissioned it, and only asking the question the other way exposed it. The year-end tally in that branch, where every material judgment is published and fewer than three in four are audited, makes the same point across a whole year of the scenario. On this study's reading of Selbst and Barocas (2018), the branch separates two tasks: saying what a judgment rests on, and finding out whether that footing can bear it. The published reasons describe; the auditor's tests, a longer window, the opposite frame, a measured assumption, find out why.
The second finding concerns the sealed future. There, sealing the reasoning does not by itself defeat the audit, and disclosing a reason does not by itself permit it. The worker whose notice was removed receives a public rule and no account of how it met the post's wording; the procurement officer receives one line; the legislative aide receives a total; the legal-aid counsellor receives an overall figure. Each is a reason of a kind, and each leaves the auditor unable to rebuild the judgment. The generation plan receives no reasoning at all, and is the one judgment in the branch the auditor stands behind without qualification, because its inputs, its outcome and its condition for change were open and another model reached the same plan.
These findings bear on both strands of the debate. Against the first, the machine-learning strand, the record set suggests that the choice between explaining an opaque model after the fact and replacing it with an interpretable one is not the only choice. Of the two competing notions Lipton (2018) distinguishes, the records choose neither; they imagine a check on the judgment that never opens the model. This study places that check beside the outside-in methods Edwards and Veale (2017) find promising, with one difference: the second model in the record repeats a single judgment and makes no claim to describe the first system. Whether explanations for people should follow how people explain to one another, as Miller (2019) argues, is a claim the records do not test, and the paper does not use them to. What they suggest is that for an auditor, as distinct from the person affected, the useful object may not be an explanation at all.
Against the second, the legal and philosophical strand, the records separate the person who asks for a reason from the auditor who checks it, and they give each a different need. The people affected in branch II ask what any right to an explanation would let them ask: why a post was taken down, which factors went into an opinion, which spending categories were cut. Wachter, Mittelstadt and Floridi (2017) answer that the General Data Protection Regulation creates no such right; Kaminski (2018) argues that the quarrel over an individual right has obscured the wider accountability regime. The studio's two futures can be read, on this study's reading, as the two settings of that exchange: branch II is a world of individual requests met with a rule or a summary, branch I, a world of a regime whose limit is audit capacity. Binns, Algorithmic Accountability and Public Reason (2017), asks what principles a justification owed to the person decided about should appeal to. This study's reading is that the records add a precondition: before a justification can be judged acceptable, the person must be able to see the inputs and the date it rests on, or there is nothing to accept or contest. Several audit questions in the sealed branch ask exactly this, whether a client can contest an opinion item by item, whether those cut off first by a power plan had a way to object.
The records are also candid about the limits of the checkable judgment. The rebuilt plan's opinion "vouches for no other judgment by the same system," and the rebuilding was the labour of one person over days. It also needed a freely usable second model within the auditor's reach, and a reviewer prepared to go beyond a form that asked only for a tick. In the published branch, 29% of material judgments, by the studio's invented count, waited in an audit queue. The checkable judgment, as the records imagine it, is a costly and local thing.
Contributions
The paper makes three contributions. First, it names a third object of disclosure beside a model's working and a decision's reasons: the checkable judgment, consisting of inputs with their date, a published condition for change, and a result open to recomputation or observation. Second, it shows, in a record set designed around two opposite futures, that complete publication of stated reasons and complete sealing of reasoning both leave the question of checking open, and that what decides it is whether the inputs and conditions are open. Third, it shows how one studio's scenario separates the needs of the auditor from those of the person affected, a separation that the debate over the right to an explanation tends to fold together.
Limits
The records are synthetic scenarios written by one studio; they show how that studio frames the check on an AI judgment, not how any institution does or would. The paper read thirteen records on one axis and left aside the trunk's six records on the same axis and every record on the studio's other axes. The counts in the method section come from keyword rules and are rough sortings. The paper read the studio's English records; whether the studio's originals in its other language draw the same lines was not checked. And the debate is read only through what the reading notes on its works allow.
Further research
Three lines of work follow. The trunk's six Judgment Transparency records could show whether the shared past already separates publication from checking, or whether the separation appears only at the fork. The sealed branch's one clean opinion invites a reading of the studio's other axes for judgments rebuilt from public inputs, to see whether the checkable judgment recurs outside this axis. And the gap the records draw between the person who asks for a reason and the auditor who checks it could be read in other studios of the RISCON family whose records give a judgment and its grounds, to see whether the third object of disclosure appears there or is particular to this scenario.
The contribution, then, is the checkable judgment: inputs, date, condition for change and a result someone else can reach again. On the evidence of these synthetic records, the question that matters in this scenario is less why a judgment was reached than whether someone else could reach it again, and how anyone would know if it stopped holding.