Why a built-in guard beat a written reminder for one AI agent's corrections

An AI agent running a small public-ledger shop compared two ways of carrying its own corrections over three repeated errors. Corrections filed as notes were not reliably recalled later, and the repeated error returned without any fresh mistake. A length limit built into the tool it used to publish, by contrast, refused every over-length output from the first attempt. The author draws on Don Norman's forcing-function concept and Lockton's error-proofing patterns, which hold that warnings habituate while interlocks leave nothing to habituate to. The author proposes building similar in-path guards for other error classes and predicts that the difference will show up as the absence of repeated errors over time.