Praxis / tool
Hate Speech Monitor (α)
A forensic, syntax-first analyzer for coordinated online harassment. Built after being the target of one. Counts structural attack patterns, not word frequency.
Methodology · 6 rules, 8 modules
Design document: 進撃の文系｜ヘイトスピーチ検出システム（α版）構築メモ
- R-01 Every post is treated as suspect by default — coverage, not selection.
- R-02 Pattern match, not frequency. Rare high-toxicity attacks are not discounted.
- R-03 One-shot patterns still count. Intent over repetition.
- R-04 Indirect subjects (she / slut / bot / pronouns) are tracked as attacks on the target.
- R-05 Structured syntax (imperatives, subject-verb-object attack frames) is prioritized.
- R-06 One post × one category = one count. No double-counting within a post.
- Attribution & quote-structure filter — victim's words quoted by attacker are excluded.
- Per-user integration across post types (text / answer / photo / video).
- Syntactic attack-template extraction (categories, not word cloud).
- Contextual emoji classification (Provocation / Sneer / Insult).
- Vision classification of image attacks (Claude Haiku 4.5).
- Statistics & visualization (this page).
- Voice-signature stylometry — hedges, softeners, imperative/self ratio, all-caps (Module 7, surfaced under each persona below).
- Cross-linguistic attack detection — Latin-dominant posts containing CJK curses, flagged per design doc note 181.
Multilingual by design. CJK curses from English-primary sources are flagged as cross-linguistic identity weaponization, not translation gaps.
Open-hostile, meme-heavy blog. Spun up within 24h of ladyyomiart's deactivation; runs a steady daily posting cadence for months.
Stylometric comparison
Per-persona category mix. The fingerprint bar shows rhetorical preference among classified posts only — Unclassified and Suspicious are excluded so the signal isn't diluted by non-attack content. Coverage % on the right shows what fraction of posts landed in a named attack category. Design doc Module 7.