Tiny prose voice concentration flat

2026-09-04, w4. Susan Warner's heroines weighted to 14.3% of supervised
words (--voice-files warner_susan --voice-weight 4, gold-weight 4, built from
ranked_convs_v2.jsonl). Melissa's verdict by ear: "decidedly not better."

The hypothesis was that the ranked pool holds 8,573 distinct (work, female
speaker) voices, so the model learns their average and an average has no voice.
Concentrating the largest single consistent voice should produce a character.
It did not — at least not at 14.3%, and not with Warner, who measured
statistically ordinary against the corpus (median reply 10 words vs 12, 59.4%
self-contained vs 57.5%, reply score 44.7 vs 45.6).

Caveat kept for honesty: there was no matched control. w4 differs from r3b
by Warner AND by gold growing 505 -> 537 AND by the re-rank. A large gain would
have survived that muddle; a null result is only weak evidence about weight 8
or about a different author.

Why: this is the seventh flat MIXTURE lever (iters, DPO, corpus slices,
base epochs, gold weight, grading rounds, voice concentration). The two levers
that ever moved her were STRUCTURAL — conversations instead of flattened pairs
(rank_conversations.py, July) and decoding config (rep penalty 1.15 + learned
</s>). Reweighting what is already in the mix keeps returning nothing.

How to apply: stop proposing mixture reweights as a fix for "she sounds
wrong." Prefer changes to what the data IS or how sampling works. See
tiny-prose-volume-is-not-the-bottleneck, tiny-prose-register-is-the-defect,
tiny-prose-sft-iters-retired, tiny-prose-decoding-config.