Six Ways to Write a Synthetic Poem

Six frontier language models — three Claudes (Opus 4.7, Sonnet 4.6, Haiku 4.5) and three GPTs (5.4, 5.4-mini, 5.4-nano) — each produced 100 poems against each of nine length- constrained prompts (5,400 calls; 5,397 returned non-null). Prompts crossed three length caps (5, 10, 20 lines) with three phrasings (write/compose/gimme). Temperature 1.0, no system prompt, max 400 output tokens. The study was preregistered with a locked coding heuristic and a 200-poem manual validation gate; deviations are logged in the public repo. Four findings emerge. (1) Family separates more sharply than tier in embedding space and on words-per-line mean, where every Claude is below 6.20 and every GPT is above 6.38. (2) Three distinct format-decoration behaviors run alongside each other on the Claude side and decouple cleanly: a markdown H1 title (Haiku-only, ~100% at every cap), a markdown bold-title (Sonnet and Opus at ~100% on cap-10/20 outputs, near-zero at cap-5), and a conversational preamble (Sonnet-only, ~100% at every cap). The cap-5 bold-title suppression is shared between Sonnet and Opus and is the cleanest within-Claude shared post-training default in the data. GPTs use none of the three. (3) Default voice does not track tier within family: GPT-nano and Claude-Opus share a lyric-I default (60% and 47% of poems); Claude-Sonnet and GPT- mini share a descriptive-third-person default. (4) A logistic regression on 13 structural features classifies poems by model identity at 56.4% (chance 16.7%) and by family at 74.3%. The largest single-feature effect is pronoun ratio (η² = 0.229), driven primarily by GPT-nano’s lyric-I behavior.