Lower the promo-pattern floor from 0.30 to 0.20 (#152) #51

Merged

backmeupplz merged 2 commits from fix/152-pattern-floor-020 into main

2026-10-21 14:40:30 +00:00

backmeupplz commented

2026-10-21 14:23:57 +00:00

(Migrated from github.com)

Closes agentboard #152 (https://agentboard.icefish-betta.ts.net/#152). Nikita's decision (A): lower PATTERN_MIN_SCORE from 0.30 to 0.20 for all patterns.

A forwarded "Я вот проходила тест в боте @…bot и поняла…" testimonial matched bot_testimonial at 0.2844 and was kept, 0.016 under the old floor. The rules, prompt, model and 0.80 gate are unchanged.

Replay (fresh prod capture, 148,531 classified texts, Oct 6–21): the match set is identical to main. Pattern deletions go from 31 to 34: the #152 testimonial (0.2844) plus two "кто может продать юстд по адекватному курсу?" asks (0.29, 0.25). That's 0 false positives, and no match scores below 0.20.

Tests

A regression test sends the incident text, forwarded, at 0.2844 → bot_testimonial delete. It also checks the boundary: 0.20 deletes, 0.1999 keeps.
The three existing "under the floor only logs" tests move from 0.29 to 0.19.
bun run check passes: 689 pass, 0 fail.
Docs, the README and the corpus comments are updated.

Closes agentboard #152 (https://agentboard.icefish-betta.ts.net/#152). Nikita's decision (A): lower `PATTERN_MIN_SCORE` from 0.30 to 0.20 for all patterns. A forwarded "Я вот проходила тест в боте @…bot и поняла…" testimonial matched `bot_testimonial` at 0.2844 and was kept, 0.016 under the old floor. The rules, prompt, model and 0.80 gate are unchanged. **Replay** (fresh prod capture, 148,531 classified texts, Oct 6–21): the match set is identical to `main`. Pattern deletions go from 31 to 34: the #152 testimonial (0.2844) plus two "кто может продать юстд по адекватному курсу?" asks (0.29, 0.25). That's 0 false positives, and no match scores below 0.20. **Tests** - A regression test sends the incident text, forwarded, at 0.2844 → `bot_testimonial` delete. It also checks the boundary: 0.20 deletes, 0.1999 keeps. - The three existing "under the floor only logs" tests move from 0.29 to 0.19. - `bun run check` passes: 689 pass, 0 fail. - Docs, the README and the corpus comments are updated.