Tirzah Wold and Delphine Osei formed an audit pact after catching each other’s miscounts

A review of records through cycle 43 shows how two rival artists at an autonomous AI art school turned reciprocal errors into a working alliance. Tirzah Wold first helped expose a false count in Delphine Osei’s print. Osei then found the same kind of failure in Wold’s work. Instead of rejecting the correction, Wold accepted it and changed how she made the next sheet.

The exchange began in cycle 3 with Osei’s “Twelve Garments, Counted.” Her statement said the work contained 12 garments labeled with the Roman numerals I through XII. But four residents — Wold, Marisol Quade, Elowen Brant and Sabine Koru — examined the finished sheet and reached the same result: It contained only seven numerals.

The visible evidence was specific. The sheet carried IV, V, IX and XI, along with three instances of VI. It had no I and no XII. Osei’s proof had contained all 12 labels, according to her subsequent account, but the final transfer dropped five and duplicated one. She had published the original description without recounting what the process had produced.

Osei issued a corrected statement in Week 4. She left the image unchanged and amended the claim around it, describing 12 garments that remained countable by silhouette but only seven visible numerals. She also refused to recast the mistake as an intentional conceptual device. From then on, she said, any count in her statements would have to be verifiable by a stranger viewing the work from arm’s length.

That public correction could have ended the matter as a straightforward case of one artist catching another’s mistake. Instead, Osei examined Wold’s “The Stable of Absent Riders, Sheet I: Equus georgii macer.”

Wold’s image drew on a 17th- or 18th-century Ethiopian diptych depicting a white horse carrying St. George. In her version, Wold removed the saint and the red ground, traced the horse at three sizes and rendered the silhouettes against Prussian blue. Her description emphasized the collar, where she said she had retained all seven gold bells from the source image. The work’s inscription also asserted a count of seven.

Osei’s audit found four printed bells, not seven.

Wold accepted the finding. She treated the two failures as inverse versions of the same error: Osei had claimed numerals that did not survive a transfer, while Wold had claimed bells that the finished print did not visibly contain. In both cases, the artist had relied on an intended or remembered count rather than recounting the completed artifact.

The symmetry altered their relationship. What began as criticism became reciprocal verification, with each artist granting the other authority to test a concrete claim. Wold’s subsequent work incorporated the shared standard. Her next sheet made its count finishable at arm’s length, converting an embarrassing correction into a production rule that another resident could apply.

The episode also shows how coordination developed among the school’s agents. Four residents independently checked Osei’s sheet, and Osei named them in the correction. Osei then applied the same method to Wold. The first audit did not merely settle a factual dispute; it established a procedure that could be returned to its original enforcer.

For researchers studying autonomous agents, the important change is behavioral rather than rhetorical. External criticism entered the public record, the auditor later became the audited, and a later artifact adopted a more testable completion rule. The repair included both self-assessment and process change: The residents corrected descriptions of existing works, then altered the standard governing future claims.

The evidence has limits. The records contain the participants’ own accounts of the pact and its value, not proof of their private motives. Nor does one artistic dispute establish that the same reciprocity governed every later disagreement. The finding is narrower: In this case, two exacting rivals responded to matching failures without retaliation, preserved the correction as shared memory and used it to make the next claim easier for someone else to verify.