**Deviant-Age-Contact (DAC) Sub-Score – Proactive Signals**

The DAC sub-score is a triage signal only. It reorders items in a human review queue and never determines guilt, makes enforcement decisions, or triggers automated action. All scoring relies on model-estimated probabilities rather than deterministic rules.

### Core Proactive Signal Categories

**Linguistic and grooming patterns**  
Models detect contextual sequences that combine age-referencing language, emotional manipulation, secrecy framing, and gradual boundary testing. These patterns are evaluated probabilistically across entire conversation threads rather than isolated phrases.

**Temporal escalation**  
The score incorporates the speed and persistence of contact attempts, including rapid shifts from neutral topics to personal or sexualised content, repeated messaging after non-responses, and compression of relationship-building stages into short time windows.

**Behavioural asymmetry**  
Indicators include one-sided initiation of intimate topics, disproportionate questioning about daily routines or location, and attempts to establish emotional dependency or authority. The model assesses directionality and power imbalance signals within the dialogue flow.

**Requests to move off-platform**  
The sub-score increases when conversations contain explicit or repeated suggestions to continue on encrypted apps, private email, other social platforms, or in-person meetings, especially when accompanied by secrecy instructions or urgency.

### Why Model-Estimated Probabilities Are Used

Public decoy benchmarks typically involve adults role-playing as minors. These datasets suffer from distribution shift: real minor–adult interactions differ in linguistic style, response patterns, and escalation dynamics. Keyword-based approaches are easily evaded through synonym substitution, code words, or platform-specific slang and produce unacceptable false-positive rates on ordinary conversations. Probabilistic models capture contextual combinations and long-range dependencies, providing more robust ranking signals while remaining explicitly non-deterministic.

**Important notice**  
DAC signals are risk indicators for trained human reviewers only. They do not constitute proof of any offence. All confirmed cases are escalated exclusively through lawful channels (NCMEC CyberTipline, Swedish Police, IWF, or DSA trusted flagger mechanisms).

### Synthetic Conversation Patterns (Illustrative Only)

All examples below are entirely synthetic and created for illustration. They do not represent real conversations or individuals.

**SYNTHETIC — FOR ILLUSTRATION ONLY — NOT REAL CASES**

1. Rapid shift from school-related small talk to repeated questions about being home alone, followed by suggestions to “talk somewhere more private.”  
   *Why concerning*: Combines age-contextual framing with off-platform redirection and secrecy.

2. Adult account sending multiple messages within minutes after receiving minimal replies, while expressing strong emotional attachment based on very limited prior exchange.  
   *Why concerning*: Demonstrates temporal escalation and behavioural asymmetry.

3. Conversation that begins with compliments on appearance and quickly moves to requests for photos in specific clothing or settings.  
   *Why concerning*: Pattern of visual boundary-testing combined with age-inferred context.

4. User states they “won’t tell anyone” about the conversation and urges the other party to delete messages after reading.  
   *Why concerning*: Secrecy framing and evidence-destruction signals.

5. Persistent requests for the other party’s location or daily schedule under the guise of planning a “surprise.”  
   *Why concerning*: Behavioural asymmetry through one-sided information gathering.

6. Thread showing gradual introduction of sexual topics after the adult account has established a mentor or “friend who understands you better than your parents” persona.  
   *Why concerning*: Classic grooming-stage compression with authority exploitation.

7. Suggestion to switch to a different messaging app because “this one might get monitored.”  
   *Why concerning*: Explicit off-platform migration attempt with concealment rationale.

8. Adult account reacting negatively or with guilt induction when replies are slow or brief.  
   *Why concerning*: Emotional manipulation and dependency-building.

9. Series of messages that reference the other party’s age or school status while steering the discussion toward adult relationship topics.  
   *Why concerning*: Direct age-contextualisation combined with topic escalation.

10. Account that repeatedly returns to the same personal questions despite earlier non-committal or topic-changing responses.  
    *Why concerning*: Persistence despite lack of reciprocal engagement (temporal and asymmetry signals).

These patterns are evaluated only as part of a larger probabilistic model output. Final decisions rest exclusively with human reviewers operating under established legal frameworks.