machine poetry · singulars

evolution

Four views of the series: where the gap with halim lives per classifier, which candidate model best matches that taste on the held-out reverse.exe, who the audience voted for at each live show, and how the machine has moved on audience-taste dimensions over time.

How the classifiers were made. Halim worked with claude opus 4.7 to read all 37 audience-decided winner and loser poem pairs from the five trained performances and surface the seven dimensions where winners consistently diverge from losers, the patterns the room reliably rewards. The methodology, the framing (audience as truth, classifier as rubric), and the calibration against contemporary poetry are halim's; claude did the systematic pass through the pairs. The rubric ends up anchored in actual audience preferences: specificity over abstraction, voice over generic lyric, earned emotional landing over decorative drift. Each classifier carries a weight of 1 to 3 reflecting how strongly it separated winners from losers.

audience truth from singulars.poems vote tallies. classifier scores from a council of 3 judges (gpt-5, claude opus 4.7, deepseek r1) against the 7-classifier rubric.