ScamScreener for Hypixel Skyblock

Model updater now treats remote models with equal or older model version as up-to-date and skips unnecessary update prompts.
Funnel detection and AI funnel context now also handle the upfront-payment funnel path (OFFER -> PAYMENT).

HOTFIX for AI Model Updater

by Pankraz01 on Feb 11, 2026

Fixed false sha256 mismatch errors when downloading AI model updates by tolerating equivalent UTF-8 payload variants (LF/CRLF, BOM/no BOM).
Fixed model release hash generation so scripts/model-version.json SHA-256 matches GitHub raw model payload bytes.

by Pankraz01 on Feb 11, 2026

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.

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 legacy training flag pipeline (TrainingFlags, TrainingTypeAiClassifier).
Removed legacy local-model migration path from root config for AI model loading.

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.

Risk warning messages still show the real player name for in-game risk message context.

by Pankraz01 on Feb 9, 2026

AI updater join-notify controls via /scamscreener ai update notify [on|off] (no args = status) and the AI update menu toggle (Join Up-to-Date Message).
Coop blacklist warning coverage expanded for co-op join request, You invited <name> to your co-op!, and <name> joined your SkyBlock Co-op!.
Coop safety flow for /coopadd <player> with [BYPASS] confirmation.
Location backbone service for scoreboard-based island detection.
SkyblockIsland enum with island names and alias matching.
Thanks to Kd_Gaming1: ModMenu integration as soft dependency with config screen support from mod list.

autoleave command behavior updated: no args now shows status; explicit toggle remains on|off.
ai update notify behavior updated: no args now shows status; explicit toggle remains on|off.
/coopadd is now always blocked first and requires explicit [BYPASS].
Mod contact metadata now defines an issues URL so the ModMenu Issues button opens https://github.com/Tangos-Mods/ScamScreener/issues/new.

Blacklist warning lookup now falls back to name-based entries when UUID lookup is unavailable.
Coop blacklist bypass message formatting now uses the shared MessageBuilder style.
ModMenu config-screen implementation fixed (correct Screen return type in the config factory).

As in previous Versions: Please provide scam messages in the issues Tracker on GitHub so I can train the model

by Pankraz01 on Feb 9, 2026

New GUI system exposed via /scamscreener settings.
Dedicated submenus for Rules, Debug, Messages, Blacklist, and AI Update.
AI update menu actions: Accept, Merge, Ignore, plus Force Check.