ScamScreener

Client-side Fabric mod for Minecraft 1.21.10 that checks Hypixel SkyBlock chat for scam risks.
The mod combines:

• a manual blacklist,
• rule-based detection (regex + behavioral signals),
• a local, trainable AI scoring system (no cloud dependency),
• and chat muting via custom patterns.

Why this mod?

Scams in trading/party situations often rely on:

• pressure ("quick", "now", "last chance"),
• upfront payment requests,
• "trust me" / "legit middleman" claims,
• external platforms/links (Discord, Telegram, /visit, etc.),
• repeated contact attempts.

ScamScreener evaluates such signals in real time and shows warnings with transparent reasons
(including rule details on hover).

Core Features

1) Blacklist Warnings

• Players can be manually added to a local blacklist.
• The mod warns when blacklisted players appear in relevant contexts:
  • team/party context,
  • party join / party tab situations.
• Warnings include name, score, reason, timestamp, and trigger context.

2) Live Chat Detection (Rules + Local AI)

Each detected player chat line is analyzed:

• Rule-based via configurable patterns (links, urgency, payment-first, etc.).
• Behavior-based via flags (e.g., "demands upfront payment").
• Local AI (logistic model with token / n-gram weights) provides an additional risk signal.
• Multi-message trend: multiple messages from the same player in a short time window can increase risk.

When thresholds are reached:

• visual risk warning in chat,
• rule list with hover explanation ("Why triggered"),
• 3x short warning sound,
• optional auto-capture into training data (from a configurable level).

3) AI Training with Local Data

• Training data is stored in CSV format.
• /scamscreener ai train trains a local model and saves it as JSON.
• Old training files/models are archived (old/ subfolder).
• Manual capturing supported: label recent messages from a specific player as scam or legit.

4) Chat Muting via Patterns

• Custom mute patterns via command.
• Case-insensitive.
• Regex supported.
• Important default logic: if no regex metacharacter is used, the pattern is matched as a
  whole word / whole phrase (not as a substring inside other words).
• Blocked messages can be periodically summarized in notifications.

Technical Overview (Simplified)

1. Minecraft delivers incoming chat/game messages.
2. The Mute Manager decides first whether a message is blocked.
3. Unblocked messages are passed to the detector:
   • parsing the player chat line,
   • rule + behavior + local AI scoring,
   • trend analysis across multiple messages.
4. On trigger:
   • warning message with level/score/rules,
   • hover with rule details + evaluated message(s),
   • optional auto-capture into training data.

Internal Architecture (Pipeline)

• Each chat line is parsed into a MessageEvent (raw + normalized).
• MuteStage filters muted messages early.
• RuleSignalStage generates signals from regex rules (patterns compiled once).
• BehaviorSignalStage generates signals from behavior (e.g., upfront payment, external platform)
  using patterns from scam-screener-rules.json.
• AiSignalStage encapsulates local AI scoring and returns a neutrally weighted signal.
• TrendSignalStage uses a per-player TrendStore (TTL deque) for multi-message patterns.
• ScoringStage combines all signals into score + level and generates explainability data.
• DecisionStage checks alert level + deduplication before warning.
• OutputStage handles UI/audio (chat warning, hover details, sound).
• Behavior patterns are defined in scam-screener-rules.json
  (externalPlatformPattern, upfrontPaymentBehaviorPattern,
  accountDataBehaviorPattern, middlemanPattern).