Answers are grounded in this paper and cite their sources.
Can a language model use a supplied specification to produce executable programs in a notation that offers an unaided reader almost no visual cues? We study emd, a stack language written in eight similar Unicode dashes, through command-line clients with tools disabled. Across 49 tasks, we collected 1150 Claude responses and an incomplete extension of 573 OpenAI responses. Exact copying is often successful, and in a complete Sol basic-prompt sample, 70 of 72 program answers produce the required behavior. Correct content does not always arrive in usable source form: output formatting rejects 33.3% of Haiku's correct answers under the basic prompt, reduced to 0.0% when the consumer's requirements are stated explicitly. We report exact-answer and normalized-content scores separately, together with token-cost estimates. These observations characterize the tested model-CLI-prompt configurations. They do not measure human readability, isolate the effect of visual opacity, or establish that the language was absent from training data.
No takes yet. Share an insight, caveat, or question.
Daniele Scaratti (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: