Tiny prose register is the defect

2026-08-30, measured on sft_r3.jsonl + ranked_convs.jsonl. Assistant
reply length by source:

Gold is the most verbose source in the mixture — more verbose than the
novels. A quarter of gold replies run 25+ words against a stated target of
1-2 sentences. It was born long (hand gold, 2026-07-20: median 17, mean 18.8);
ouroboros adds only +1 word, so this is a founding property, not loop drift.

Every selection mechanism aims above target: score_pair's length term
(30 pts, the heaviest, band 8-30 words), substance (24 pts, rewards
non-stopword content over pleasantries), the grading UI (short replies read as
thin), and gold weight. Plays are the only short-turn source and score 15
points LOWER — penalised precisely for being conversational. --play-weight
was cut to 1 on 2026-08-25 for volume bookkeeping (book corpus grew 47%); no
one checked register.

This explains the gold-weight trade that was previously unexplained: 11% -> 25%
gold bought character and cost curiosity because it upweighted the longest,
most expository material in the set.

Plays never helped for two separate reasons: the ranker penalises them, AND
play dialogue is SCENE dialogue — right form (strict alternation, no narration),
wrong self-containedness (every line points at offstage context).

How to apply: the untested lever is retargeting selection, not adding data
— narrow score_pair's length band toward the target, raise --play-weight,
re-rank, rebuild, retrain. Zero new grading. See
tiny-prose-volume-is-not-the-bottleneck, tiny-prose-goals-constraints.