Safety Is Separate From Scoring

A high score and a safety-sensitive input are not the same condition. Intense political criticism may contain substantial harm language without expressing immediate personal danger. A short sentence can contain a direct self-harm or violence signal even if the ordinary scoring components would be incomplete. Conflating these cases would make both the score and the response less responsible.

The analysis engine therefore runs a separate safety detector after the deterministic score. It checks narrowly defined patterns across five categories:

• self-harm;
• direct violence;
• graphic harm;
• harassment;
• medical distress.

The result is none, caution, or sensitive, together with matched patterns and bounded guidance. Any match activates the safe-reply path. Self-harm, direct violence, graphic harm, and medical-distress matches are treated as sensitive; harassment is treated as caution. The response builder avoids playful score language and does not amplify harmful intent.

This separation also explains why somatic language can affect the score without automatically activating the safety route. “My heart is racing” is a somatic-tension signal. “Chest pain” or “panic attack” is also covered by the medical-distress safety detector. Those rules are intentionally explicit and inspectable.

Pattern matching remains limited. It can miss euphemism, irony, multilingual language, obfuscation, or context outside the submitted text. It can also match quoted or educational content without understanding why it was included. The safe-reply path reduces one class of response risk; it is not a complete moderation, clinical, or crisis system.

For that reason, the checker must never be positioned as emergency support. Safety-sensitive output should remain restrained, avoid medical advice, and encourage immediate human or professional support when the language indicates urgent risk.