The fifteen synthetic records on ASI 2030's Concentration axis cannot show what concentration in AI will do. They show how an invented record set built around a human auditor reasons when the judges come from a few providers, and, read as one line through both imagined futures, they answer consistently: an auditor does not accept agreement as a check until the sources that agree are shown to be independent, and the survival of a second source decides what an audit can still do.
Discussion
Set against the work on monoculture, the records keep the worry and move its object. Kleinberg and Raghavan (2021) and the studies of homogenized outcomes by Bommasani, Creel, Kumar, Jurafsky and Liang (2022) and by Toups and colleagues (2023) are about the receiving end: welfare, exclusion, the same applicant turned away everywhere, and Creel and Hellman (2022) explain why that matters. The synthetic records share the fear of failing together, in the bond and forecasting cases of the trunk and in the festival labels of branch I. But their auditors put a prior question: what is the agreement of two models worth as evidence? The re-rating record answers it most sharply, reading a low disagreement rate as a possible sign of herding rather than its refutation. Read this way, a monoculture is first an epistemic condition for whoever must check a judgment, and only afterwards a pattern in who is hurt.
Set against the work on concentration, the records narrow a broad agenda to one practical stake. Vipra and Korinek (2023) write about contestable markets and quality standards, Sastry and colleagues (2024) about compute as a lever, Widder, Whittaker and West (2024) about the limits of openness, and Plantin, Lagoze, Edwards and Sandvig (2016) about platforms turned infrastructure. This paper reads each theme as having a counterpart somewhere on the axis: the export restriction, the single-provider contract, the open-weight model under a share-based formula, the grid run on one provider's system. What the auditors ask of each, though, is whether anything remains to compare the judgment with. In branch II nothing does, and the record set writes the consequence into its reopening lines. A judgment there is reopened not when a figure crosses a threshold but "once a second, cross-checkable provider is operating again." The structure of the market has become a condition of audit.
Two further points follow. First, of the kinds of question counted, the auditors' most frequent is about proximity rather than size: whether the beneficiary supplies the data, the recommending tool belongs to the provider it recommends, or the vendor reports on its own performance. Concentration in these records is less a matter of how much one firm holds than of how little distance is left between judging, supplying the grounds and gaining. Second, the people judged appear less often than the sources. Only 4 of the fifteen records have an audit question asking whether households, organizers, renters or excluded consortia have a voice or a way to object. The monoculture literature's focus on those who bear the outcome is, in this thread, more often implied than asked.
Contributions
The paper makes three contributions. It reads ASI 2030's Concentration axis whole, across the studio's fork, and shows that it is organized around a single question, whether a judgment has an independent second source, which the trunk poses, branch II loses and branch I, in between, redistributes. It connects the debate on algorithmic monoculture and the debate on concentration in AI's markets and infrastructure through a practice that, in this paper's reading, neither takes as its centre, the checking of a judgment by comparison, and proposes that concentration be read, among its other effects, as the removal of that comparison. And it shows, in a record set designed for audit, how the standard of evidence an auditor applies, agreement counts only between independent sources, turns the structure of a market into a condition on whether an opinion can be given at all.
Limits
The limits are those of the evidence. Every record is synthetic: the providers, blocs, model families, crises and figures are invented, and nothing in this paper is a claim about any real system, market or future, or a forecast about artificial general or superintelligent AI. The fifteen records are one thread chosen by one rule; the library's other axes were not read. The reading of the trader's two choices as the studio's two futures rests on matching labels and period titles and is this paper's interpretation. The counts are small and descriptive, and the frame was measured only among the fifteen records, not across the studio's whole library. Where the works are used as lenses, the readings that go past their stated arguments are this paper's own; in particular, Widder, Whittaker and West's argument about openness fits the audit-hours record only partly. The paper recommends no policy, procurement, trading or energy decision, and does not say which opinion an auditor should give.
Further research
Three lines follow. One is to read the studio's other axes for the same standard, to see whether the demand for an independent second source holds where concentration is not the named theme. Another is to compare this synthetic auditor's reasoning with other studios' record sets in the same family. The third is conceptual: the monoculture literature measures homogeneous outcomes and the concentration literature attends to market power, but a measure of how many genuinely separate sources remain to check a judgment against would join them. The records read here suggest why such a measure matters: when the same model is asked twice, the second answer is not a second opinion.