Tiny prose character is topic
2026-08-25. Melissa, after chatting: "v10 is measurably worse but the
individual replies are more interesting. I liked how the old Jane would bring
up ghosts and got off on tangents."
Two causes. Check the sampling settings FIRST. chat.py had been lowered
to temp 0.6 to stop runaways, and that alone turned "Never. My circle of
friends is limited." into "No." Sampling parameters are part of the
evaluation, not a detail — changing them invalidates every judgement about
character made afterwards. Back at 0.8 / 100 / 120.
The real cause is topical. prose/scripts/topics.py buckets conversations
by SUBJECT (where coverage.py covers conversational FUNCTION). The ranked
pool is money 17.4%, love 15.0%, gothic 0.7%, and 60% match nothing.
Two plausible culprits were tested and are WRONG: score_pair's OOV penalty
does not remove it (top 8,000 by score is more gothic than the bottom), and
the dialect exclusion only accounts for ~24 conversations. It was never there:
the Gutenberg authors are domestic realists — Trollope, Yonge, Oliphant,
Gaskell. Ghosts are not their business.
Fix is mining blbooks_sft for it, same as the greetings
(tiny-prose-dialogue-extraction-bug). Two things make that work:
CORROBORATE— require two DISTINCT gothic terms across the window.
Victorian fiction is full of "a ghost of a smile" and "haunted by doubt";
one hit means metaphor, repetition means subject.
- Trim to the clean run, don't drop the window.
looks_ocr_damagedwith
any() over the window made wider --context reduce yield. Trimming
doubled everything.
Full text of the analysis: prose/notes/v12-sweep.md.