Tiny prose volume is not the bottleneck

2026-08-30. Chatterbox's SFT mix (~42,648 items): ~15k rewritten QA, ~9k
synthetic conversations, ~8k corpus-grounded, ~8k extended QA, ~2k
greetings/edge cases/factual QA, 648 play scenes. Roughly 23% multi-turn,
~75% machine-written or machine-rewritten. Melissa: it "is probably worse than
Jane at holding a conversation."

Jane, unweighted, from sft_r3.jsonl: 11,329 distinct items, ~77%
multi-turn (ranked 6,768 @ 11.3 turns, gold 1,677 @ 5.9, graded 294).
3.8x less data, 3x the conversational shape, better thread-holding.

Why: this retires "the bottleneck is distinct supervised turns" — the
premise behind DPO round 1, the iters sweep, the plays-only/attributed-only
slices, the ouroboros loop and the grading grind. Five experiments, all flat,
all reasoning from a premise the control disproves. It also shows the
no-synthetic rule costs nothing: chatterbox is ~75% synthetic and lost.

Thread-holding was solved in July by rank_conversations.py replacing
rank_pairs.py (conversations, not flattened pairs). What remains is
REGISTER — see tiny-prose-register-is-the-defect.

How to apply: do not propose more grading, more gold, or more mining as a
fix for "she sounds wrong." Volume is answered. Supersedes the closing claim in
tiny-prose-sft-iters-retired. See also
tiny-prose-goals-constraints, tiny-prose-dpo-round1.