  ElebobaandClaude Opus 5.5
|
97f3173c9f
|
feat(core): safety detector with RU risk lexicon and fixed crisis script
- check(text, llm_signal) -> SafetyResult: normalisation (ё, homoglyphs, translit,
repeats), stem regexes per category, negation window (acute -> possible only),
weapon+risk combination, final level = max(lexicon, llm)
- crisis_reply(categories, audience) loads config/texts/crisis.yaml (patient/family,
domestic violence and harm-to-others variants)
- 99-phrase eval set; acute recall 100%, < 5 ms per message
Refs #4
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
2026-10-03 15:01:35 +03:00 |
|