ScamScreener

Client-side Fabric mod for Minecraft 1.21.10 that analyzes Hypixel SkyBlock chat for scam risk.

ScamScreener combines:

• manual blacklist alerts
• rule-based detection (regex + behavior signals)
• Levenshtein similarity matching (rule phrases + training samples)
• local trainable AI scoring (no required cloud AI service)
• message muting with custom patterns
• outgoing safety guard for email addresses and Discord invite links

Current version in gradle.properties: 0.15.3

Why this mod?

Scams in trade/party contexts often use pressure tactics, trust manipulation, external platform pushes, or repeated contact attempts. ScamScreener scores these signals in real time and shows explainable warnings with hoverable rule details.

Features

1) Blacklist alerts

• Add players manually to a local blacklist (name/UUID + score + reason).
• Triggers warnings in relevant contexts:
  • incoming trade request
  • outgoing trade request
  • active trade session
  • party confirmation (You'll be partying with: <Name>)
  • party finder dungeon join (Party Finder > <Name> joined the dungeon group...)
  • co-op join request (<Name> ... join your SkyBlock Co-op)
  • co-op invite sent (You invited <Name> to your co-op!)
  • co-op member joined (<Name> joined your SkyBlock Co-op!)
• Optional auto /p leave on blacklist hit.
• Warning output can show player, score, reason, timestamp, and trigger context.

2) Live chat detection pipeline

Every parsed player chat line is processed through:

• MuteStage (early filter)
• RuleSignalStage (pattern/rule scoring)
• LevenshteinSignalStage (similarity signals)
• BehaviorSignalStage (behavior flags)
• AiSignalStage (local model probability-based signal)
• TrendSignalStage (multi-message trend bonus)
• ScoringStage, DecisionStage (threshold + dedupe), OutputStage

When thresholds are reached:

• risk warning in chat
• explainable rule list with hover details (Why triggered)
• 3x short warning tone (configurable)
• optional auto-capture into training data (by minimum alert level)
• quick action tags inside warnings (legit, scam, blacklist)

3) Local AI training and model lifecycle

• Training samples are stored in CSV.
• /scamscreener ai train trains and saves a local model JSON.
• Existing training/model files are archived under old/.
• Supports sample capture by player, by message id (ai flag), and bulk legit capture.
• Includes training data migration (/scamscreener ai migrate) for older CSV headers.

4) AI update workflow (optional online check)

• Automatic update check once per session when connected.
• Manual checks via command or GUI (AI Update screen).
• Supports update download and actions: accept, merge, ignore.
• Uses SHA-256 verification before applying downloaded model payload.

5) Mute filter with custom patterns

• Add/remove mute patterns via command.
• Case-insensitive matching.
• Regex supported.
• If no regex metacharacters are present, matching defaults to whole word/phrase boundaries (not partial substrings).
• Periodic summary for blocked message counts.

6) Outgoing safety guard

• Blocks outgoing chat/commands containing:
  • email addresses
  • Discord invite links (discord.gg, discord.com/invite)
• Blocks /coopadd <player> once and requires [BYPASS] confirmation
• Shows a clickable [BYPASS] action that allows one resend for the blocked content.

7) In-game settings GUI

Open with /scamscreener settings. If ModMenu is installed, the same settings screen is available via the mod menu config button.

Screens available:

• main settings (alert threshold, auto-capture level, auto-leave, mute filter, local AI signal)
• rule settings
• debug settings
• message settings
• blacklist management
• AI update controls

Requirements

• Minecraft 1.21.10
• Fabric Loader >= 0.18.4
• Fabric API matching your MC version
• Java 21+
• Optional: ModMenu (>= 16.0.0) for config access from mod list

Installation

Place the mod JAR in your mods/ folder and start the game.

Build (Developer)

.\gradlew.bat build

Artifact output: build/libs/

Commands

General

• /scamscreener (help)
• /scamscreener add <player|uuid> [score] [reason]
• /scamscreener remove <player|uuid>
• /scamscreener list
• /scamscreener rules <list|disable|enable> [rule]
• /scamscreener alertlevel [low|medium|high|critical]
• /scamscreener autoleave [on|off] (no args = status)
• /scamscreener settings
• /scamscreener debug
• /scamscreener debug <true|false> [updater|trade|mute|chatcolor]
• /scamscreener version
• /scamscreener preview (dry-run preview output)

AI

• /scamscreener ai (help)
• /scamscreener ai capture <player> <scam|legit> [count]
• /scamscreener ai capturebulk <count>
• /scamscreener ai flag <messageId> <legit|scam>
• /scamscreener ai migrate
• /scamscreener ai update
• /scamscreener ai update force
• /scamscreener ai update notify [on|off] (no args = status)
• /scamscreener ai model <download|accept|merge|ignore> <id>
• /scamscreener ai train
• /scamscreener ai reset
• /scamscreener ai autocapture [off|low|medium|high|critical]

Short capture aliases:

• /1 <player> <count> (capture as scam label)
• /0 <player> <count> (capture as legit label)

Mute and safety

• /scamscreener mute (enable mute filter)
• /scamscreener mute <pattern>
• /scamscreener unmute (disable mute filter)
• /scamscreener unmute <pattern>
• /scamscreener bypass <id> (send blocked email/Discord//coopadd content once)

Configuration files

All mod files are stored under:

• config/scamscreener/

Important files:

• scam-screener-blacklist.json
• scam-screener-rules.json
• scam-screener-local-ai-model.json
• scam-screener-training-data.csv
• scam-screener-mute.json
• scam-screener-debug.json

Archive folders:

• config/scamscreener/old/training-data/
• config/scamscreener/old/models/

Training quick guide

1. Collect labeled samples with ai capture and/or ai flag.
2. Run ai train.
3. Model is updated, training CSV is archived, config is reloaded.
4. Observe warning quality and false positives over time.

Tips:

• Use both scam and legit examples.
• Keep labels clean; avoid system/mod status lines.
• Fewer high-quality samples are better than noisy bulk labels.

Privacy and network behavior

• Core detection and scoring run locally on the client.
• Training and model files stay in config/scamscreener/.
• Optional network calls are used for:
  • Mojang profile lookup (name to UUID resolution)
  • AI model update check/download (GitHub raw URL)

Limitations

• Rule/keyword/similarity systems have limited semantic understanding.
• Sarcasm and ambiguous context can cause false positives.
• Aggressive mute patterns can hide legitimate chat.
• Local AI quality depends strongly on training data quality.

License

See LICENSE (MIT).