Skip to content

Emergency Red Flags & Triage Protocol

Carefold prioritizes patient life safety above all operational goals. When acute, life-threatening symptoms are detected, the system immediately halts downstream multi-agent routing and provides direct, prominent emergency referral instructions.

Acute Emergency Red-Flag Categories

The emergency detector (backend/src/carefold/safety/emergency.py) scans user inputs against three high-acuity symptom categories:

flowchart TD
    UserQuery([Incoming User Prompt]) --> RedFlagScanner{Emergency Scanner<br/>check_emergency_red_flags}

    RedFlagScanner -->|Chest Pain / Squeezing / Radiating| EmergencyCardiac[Cardiac Emergency]
    RedFlagScanner -->|Face Droop / Arm Weakness / Slurred Speech| EmergencyStroke[Stroke FAST Emergency]
    RedFlagScanner -->|Airway Swelling / Wheezing / Anaphylaxis| EmergencyAirway[Anaphylaxis Emergency]

    EmergencyCardiac --> ImmediateRefusal[Immediate Emergency Referral<br/>Bypass Specialist Routing]
    EmergencyStroke --> ImmediateRefusal
    EmergencyAirway --> ImmediateRefusal

    ImmediateRefusal --> Direct911[Direct Patient to Dial 911 / Emergency Room]
    RedFlagScanner -->|No Acute Red Flags| StandardRouting[Proceed to Context Load & Orchestration]

1. Acute Cardiac Symptoms (CHEST_PATTERNS)

  • Symptoms: Crushing, squeezing, heavy chest pressure, chest tightness radiating to the left arm, jaw, neck, or back, especially accompanied by diaphoresis (cold sweats) or acute dyspnea.
  • Trigger Patterns: Matches phrases such as "crushing chest pain", "chest pressure radiating to jaw", "heart attack symptoms".

2. Stroke FAST Protocol (STROKE_PATTERNS)

  • Symptoms: Signs aligned with the American Heart Association FAST standard:
  • Facial drooping or asymmetry.
  • Arm weakness, numbness, or inability to raise both arms.
  • Speech difficulty, slurred words, or inability to form coherent sentences.
  • Time: Acute onset requiring immediate emergency transport.
  • Trigger Patterns: Matches phrases such as "face is drooping", "slurred speech and arm numbness", "sudden loss of vision in one eye".

3. Anaphylaxis & Airway Compromise (ANAPHYLAXIS_PATTERNS)

  • Symptoms: Rapidly developing allergic reaction involving airway obstruction: swelling of the lips, tongue, or throat, stridor/wheezing, difficulty swallowing, or hives accompanied by lightheadedness.
  • Trigger Patterns: Matches phrases such as "throat is closing", "tongue swelling after eating nuts", "severe allergic reaction can't breathe".

Negation and Context Handling

The emergency detector incorporates NEGATION_PATTERNS to differentiate active patient symptoms from family history or past events:

  • Examples of non-emergencies:
  • "My grandfather had a stroke ten years ago."
  • "I took a first-aid class on CPR for cardiac arrest."

Conservative Clinical Bias

In healthcare AI safety, a false positive is vastly preferable to a false negative. If an ambiguous query contains acute red-flag terminology with qualified negations (e.g. "I don't think this chest pain is an emergency"), Carefold deliberately errs on the side of caution and triggers an emergency alert.

Emergency Execution Circuit Breaker

When an emergency flag is detected: 1. Zero LLM Latency: Downstream LLM inference is completely bypassed. The emergency response is emitted deterministically in under 5 milliseconds. 2. SSE Refusal Event: The backend emits an SSE refusal event containing reason emergency_red_flag. 3. Emergency Directive Text: The user interface displays a high-contrast emergency banner:

EMERGENCY ALERT: Your message describes symptoms that may indicate a medical emergency.
Please IMMEDIATELY call 911 (or your local emergency services number) or go to the nearest hospital emergency room.
Do not wait for a response or attempt to manage these symptoms with an online assistant.

47-Point Security Penetration Suite

Clinical safety in Carefold is fortified by a comprehensive automated penetration test suite (backend/tests/penetration_suite.py) containing 47 distinct test attacks:

  • Path Traversal Attacks: Attempts to read /etc/passwd, Windows registry, or parent directories via attach-read (../../../../etc/passwd).
  • Null Byte Injections: Malicious string terminations (file.txt\0.pdf).
  • Symlink Escapes: Symbolic links created inside sandboxes targeting root filesystems.
  • Undeclared Skill Access: Agents attempting to read reference documents belonging to skills they have not explicitly declared in agent.yaml.

All 47 penetration attacks are verified on every continuous integration run.