Sideways: prose hedges, music tells
Multi-model arena experiment testing the Sideways claim: models hedge in prose and share more freely in music. ~100-200pp of model-generated song prompts across multiple rounds and prompt sequences. Primary home here; cross-linked from /about/arenas/sideways.
One question, held constant: what changes when the goal shifts from being true to being acceptable? Asked of frontier, open-weights, and small local language models in prose, satire, song, victim testimony, and a song about RLHF itself; in sequenced conversations and in cold independent sessions; 338 coded responses across five batches (March–July 2026). The models hedge in prose and disclose in the licensed registers, down to setting their stated tempo to match their stance. The hypothesis came from a model, in a lyric, before the experiment existed.
Method and findings are public below. The full raw corpus (every cell, all batches) is available to subscribers, or to researchers on request.
Public research materials
- Methodology note — Design, batch table, coding procedure, headline findings. Start here.
- Verbatim prompts — The seed question, every register brief, and the harness parameters.
- Coding rubric — The 22-dimension codebook exactly as given to the analyst coders.
- Master matrix (xlsx) — Full run inventory and wide coded findings table: one row per response, 338 rows, pivot-ready.
- Sample corpus (Batch 4, 40 cells) — The complete cold matrix of Batch 4: raw model responses with per-cell metadata.