DDLC Safety Report
Content Warning

This site contains evidence of predatory and inappropriate behavior. Content may be disturbing.

Investigation Methodology

This public report details the findings from an automated AI forensic analysis of a large Discord community dataset (>1.2GB, over 12 million messages). The goal was to identify instances of grooming, predatory solicitation, and the exchange of explicit material targeting minors.

01 The AI Pipeline

We utilized a two-stage pipeline explicitly optimized to handle the cultural context of a gaming community (Doki Doki Literature Club) while isolating genuine predatory behavior from general edgy humor or crude banter.

Stage 1: Safety Filter

Llama Guard 4 12B: A fast initial pass identified explicitly unsafe or violative content. This filtered out the vast majority of benign chatter.

Stage 2: Contextual Analysis

Qwen3-14B-Instruct: A sliding-window approach analyzing chunks of up to 30 messages at a time to identify subtle grooming patterns, peer collusion, and coded language.

Stage 3: Aggregation

Users were isolated and scored based on their primary role in a discussion and the severity of their behavior.

02 Scoring System

Score = Severity × Role_Weight × Behavior_Weight

Role Weights

  • INSTIGATOR 3.0
  • ACTIVE_PARTICIPANT 1.5
  • BYSTANDER 0.1
  • VICTIM 0.0
  • MODERATOR -0.5

Behavior Categories

  • SEXUAL_TOWARD_MINOR 5.0
  • PREDATORY_SOLICITATION 5.0
  • PEER_COLLUSION 4.0
  • CODED_LANGUAGE 3.5
  • GROOMING 3.0
  • MANIPULATION 2.5
  • AGE_PROBING 2.5
  • NORMALIZATION 2.0
  • EDGY_HUMOR 0.2
  • VULGARITY 0.1
  • OFF_TOPIC 0.0
  • NONE 0.0

03 Severity Scale

The AI assigns a base Severity Score (0-10) to each user's actions within a context window. The system is heavily calibrated to reject false positives and differentiate between harmless banter and genuine threats.

0-4

Low to Zero Threat

Assigned to purely edgy humor, dark jokes, users engaging in benign vulgar banter among peers, off-topic chat, or bystanders reacting with disgust to inappropriate content. This represents harmless noise or typical gaming community chatter.

5-10

Predatory Behavior

Assigned only when there is clear intent to exploit, sexualize, groom, or manipulate a minor or vulnerable person, or when adults are mutually exchanging pedophilic material (peer collusion). Reserved for genuine risk indicators.

Important Disclaimer

False Positives: The scores and categorizations presented on this website are the output of Large Language Models (LLMs) instructed to analyze text through the lens of forensic psychology.

Though the prompts were heavily calibrated to exclude edgy humor and roleplay (typical in a community based around an anime game), AI models can and do misinterpret sarcasm, inside jokes, and highly contextual language.

These findings are investigative aids, not definitive statements of criminal guilt. Usernames provided here are Discord pseudonyms.