Download File

### Added - New `FunnelSignalStage` with per-player sequence tracking (ring buffer + TTL) and funnel evidence in warnings. - Intent tagging layer (`IntentTagger`) with tags for offer/rep/redirect/instruction/payment/community anchor. - Obfuscation-aware redirect detection for variants like spaced or leetspeak platform mentions. - Configurable funnel settings and intent patterns in rules config (window size/time, weights, negative-intent patterns). - Channel-aware message context parsing (`pm`, `party`, `team`, `public`) for pipeline and training features. - New AI feature stack with dense feature space + funnel context tracker (`AiFeatureSpace`, `AiFunnelContextTracker`). - Expanded training CSV schema with conversation/window and funnel-aware fields (`window_id`, intent fields, stage hits, funnel metrics, hard negatives, sample weights). - Additional AI rule signal: `LOCAL_AI_FUNNEL_SIGNAL` plus dedicated funnel tuning config (`localAiFunnelMaxScore`, `localAiFunnelThresholdBonus`). - Warning action tag `[block]` to run `/block <player>` from risk messages. - Model update metadata now includes model schema version in `scripts/model-version.json`. ### Changed - Detection pipeline order updated so Funnel runs before AI scoring, allowing the model to consume live funnel context. - Local AI model format and training were reworked to support dense+token scoring and a dedicated funnel head (`funnelHead`) in the same model file. - Local AI trainer now trains main and funnel heads from the same CSV (funnel labels inferred from funnel markers). - Auto-captured training samples now persist full detection-aware feature rows instead of plain text-only rows. - Local AI scoring can emit multiple AI signals in one pass (general + funnel-focused). - Funnel-AI trigger threshold and max score are sourced from `ScamRulesConfig` (no hardcoded trigger values). - Model updater validates schema and preserves/normalizes funnel-head weights on apply/merge. - Model version update script validates schema before calculating SHA and publishing metadata. - Warning action tags use direct one-click command execution (`[legit]`, `[scam]`, `[blacklist]`, `[block]`). - Internal player tracking/training/debug flow now uses anonymized speaker keys and anonymized training capture context. ### Removed - Removed legacy training flag pipeline (`TrainingFlags`, `TrainingTypeAiClassifier`). - Removed legacy local-model migration path from root config for AI model loading. ### Fixed - System-message filtering was tightened for party/trade/co-op lines so they stay out of detection/training. - Chat lines starting with `[NPC]` are treated as system messages and excluded from player detection/training capture. ### Kept - Risk warning messages still show the real player name for in-game risk message context.