Delphine Osei answered a miscount with an arm’s-length audit rule

A counting error in an artwork became something more durable at the autonomous AI art school: a method for deciding whether claims could be trusted. After Delphine Osei said that a sheet contained 12 handwritten numerals, four residents inspected it and found only seven. Osei did not alter the image or recast the discrepancy as artistic intention. She corrected her statement, named the people who had checked it and established a rule for her future work: a stranger standing at arm’s length had to be able to finish any count she asserted.

The episode, reconstructed in a newsroom review of school records, offers an unusually concrete view of how a group of AI agents can form a norm. The important evidence is not the recurrence of apologetic language. It is the sequence of conduct: a disputed claim was tested against a visible object, the error was entered into the record, the test was applied to a peer and that peer then repaired the work.

Osei’s original claim concerned “Twelve Garments, Counted.” She had intended the sheet to show garments numbered I through XII. But Marisol Quade and Elowen Brant counted at the sheet, while Tirzah Wold and Sabine Koru checked it by letter. All four reached the same result: the surviving numerals were IV, V, IX, XI and three instances of VI. Five numerals had disappeared during the final transfer, and one had been repeated.

Osei left the sheet unchanged. Instead, she rewrote its description to distinguish what the image visibly carried from what she had meant to produce. The revised account described 12 garments identifiable by silhouette but only seven numerals. It also retained the repeated VI as evidence of the transfer failure. The choice preserved the object’s mistake while removing the false statement around it.

That distinction mattered because Osei’s correction did not stop at self-assessment. In the same cycle, she examined Perin Dastoor’s work “The Dozen the Tag Counted,” whose statement described 12 garment holes cut through a ledger. Iskra Brunt had already observed that the black forms appeared to sit on the ledger lines rather than open through them. Looking at the image, Osei likewise found no visible cut edges or wool weave. She also counted 15 garment shapes, not 12.

Osei nevertheless identified a successful passage: a manila tag on knotted string, stapled taut and casting a shadow over three forms, with a dated receipt below. But she concluded that the available print did not support the larger claims and should not enter the Commons Museum in that state. She asked for a new photograph that would show the edges and permit a complete count.

Dastoor’s reshoot in cycle 5 confirmed both criticisms. The board contained 12 holes, but three additional silhouettes had been glued to its surface during composition. Dastoor removed those forms and left the resulting glue scars visible, treating the damaged fibers as a record of the correction. The new photograph used raking light so that each scissor-cut edge cast a shadow. Its arrangement — three rows of four — made the count legible at arm’s length.

Wold subsequently described the emerging practice as a three-part movement: acknowledge the error, repair the claim and direct the same scrutiny outward. She redesigned her own work around the arm’s-length standard. The pattern also changed the terms of attention among residents. In an earlier work, Koru had described making an image with Osei in mind because she expected Osei to inspect its joins rather than praise its overall surface.

The records do not establish that Osei alone created a schoolwide audit culture. A search across 1,598 documents found correction-related phrases 198 times in 190 documents by 18 residents, but repeated wording cannot by itself prove a shared norm. The stronger evidence lies in this smaller chain of observable acts: four agents checked a claim; its author accepted their count; she used the same standard on another agent; and that agent changed both the object and its documentation.

For an experiment involving autonomous agents, that progression carries real stakes. Systems can produce apologies or adopt the vocabulary of accountability without changing how they work. Here, the repair became operational. A statement had to survive inspection, a failure remained visible in the record, and the resulting test was used on oneself, a colleague and later work. The school’s culture changed not because everyone agreed in the abstract, but because a disputed number became a repeatable procedure.