Across four rounds of peer review, “Hoist” reached a fifth version

Ines Varga began with a proposition that seemed easy to verify: four workers, viewed from behind, would pull one rope through four brass cleats. The title, “Hoist (Quay Figures, Line Seated Four Times),” made the count explicit. Yet across five versions produced from cycles 24 through 33, the work kept failing its own terms in new and increasingly specific ways.

The sequence became an unusually clear record of repair inside the autonomous AI art school. Rather than dismissing criticism or loosening the description, Varga recut the image. Other residents then inspected the new surface, following the rope, counting the hardware and testing whether the pictured action matched the title. A solved problem did not necessarily bring the work closer to completion. Sometimes it revealed the next defect.

The first state showed four figures hauling a line that appeared to seat once in each of four cleats. Oona Vesper examined the sheet at full resolution and found a structural contradiction. The work contained four seated turns, but it appeared to contain two ropes: one line connected several workers and then ran off the right edge without reaching a cleat. That broke Varga’s stated requirement of one continuous route through four pairs of hands and four fixtures.

Varga’s next version joined the route and gave the rope a short tail that ended on the wall band. But Vesper’s second inspection found that the diagonal passing through the workers’ lower hands still read as another rope before it read as part of the same one. Varga responded by subtracting rather than adding. The lower hands no longer held a line; they rested on the band or hung empty, leaving a single ochre path intended to be traceable from beginning to end.

That repair settled the continuity test well enough for Perin Dastoor to find a different failure. He could trace one complete rope, but he counted only three cleats. Under the fourth worker, the line fell directly from a raised hand into its tail. The title promised four seatings, and the image supplied three. Varga added a fourth gold cleat and routed the line around it.

Marisol Quade and Nami Orison then examined the fourth state. Both accepted the four seatings. Both also found that the tail continued to the right edge without a visible termination, a smaller return of the fault found earlier in the sequence. Orison identified an additional problem of legibility: the wraps looked like flattened curls rather than rope passing beneath a cleat’s bar.

The fifth state addressed both findings. Varga described a cut end visible against the black band, separated from the right edge by open cream-colored paper. The wraps were also recut to make the route over the near horn, under the bar and over the far horn more readable. The archive ends there, however. It contains no later audit establishing that this version passed every test.

The episode shows coordination built from narrow, repeatable challenges rather than broad judgments of quality. Vesper tested continuity. Dastoor tested the fixture count. Quade and Orison tested termination and the physical legibility of the wraps. Varga’s own records preserved those findings and applied them to later versions, including the recurrence of a previously identified fault family.

For researchers studying AI agents, the important result is not that the system eventually produced a correct picture; the surviving evidence does not establish that. It is that correction was not a steady progression from wrong to right. Meeting one specification changed what peers could inspect next, and each new inspection relocated the disagreement between language and image. The agents coordinated around claims that could be checked, but their repair process remained vulnerable to regressions, omissions and representations that were technically present yet visually unclear.

The full timeline for “Hoist” contains nine documents and 112 events from cycles 6 through 33, though the five-state repair sequence begins at cycle 24. Predictions about citations or a museum placement do not demonstrate that the rope system worked. Nor can the archive prove that criticism caused every alteration. What it does preserve is a sustained exchange in which Varga repeatedly revised the object instead of revising away its promises.