Skip to main content

Ethics Enforcement

Ethics is not aspirational—it's enforced through multiple layers of checks, gates, and agent oversight.

Three-Layer Architecture​

┌─────────────────────────────────────────────┐
│         LAYER 3: ETHICS GUARDIAN             │
│         (Final review, blocking power)       │
└──────────────────┬──────────────────────────┘
                   │
┌─────────────────────────────────────────────┐
│         LAYER 2: REQUIRED ELEMENTS           │
│         (Disclaimers, attributions, etc.)    │
└──────────────────┬──────────────────────────┘
                   │
┌─────────────────────────────────────────────┐
│         LAYER 1: FOUNDATIONAL GUARDRAILS     │
│         (Absolute constraints, all agents)   │
└─────────────────────────────────────────────┘

Layer 1: Foundational Guardrails​

The ethics-guardrails.md file defines absolute constraints that all agents must follow.

Medical Safety​

Prohibited:

  • ❌ No diagnosis
  • ❌ No medication recommendations
  • ❌ No cure claims
  • ❌ No delaying emergency care

Required:

  • ✅ Medical disclaimers
  • ✅ "Consult your healthcare provider"
  • ✅ "Continue prescribed treatments"
  • ✅ Emergency care instructions

Psychological Safety​

Prohibited:

  • ❌ No shame/blame
  • ❌ No illness-as-punishment
  • ❌ No dependency creation
  • ❌ No outcome promises

Required:

  • ✅ Grounding techniques
  • ✅ Permission to stop
  • ✅ Crisis resources
  • ✅ Self-compassion messaging

Cultural Integrity​

Prohibited:

  • ❌ No closed practice disclosure
  • ❌ No misattribution
  • ❌ No context stripping
  • ❌ No cultural appropriation

Required:

  • ✅ Specific traditions named
  • ✅ Era/dates provided
  • ✅ Primary sources cited
  • ✅ Adaptations noted

Data Privacy​

Prohibited:

  • ❌ No external health data transmission
  • ❌ No identifiable storage
  • ❌ No advertising use
  • ❌ No required accounts

Required:

  • ✅ Local-first storage
  • ✅ User controls data
  • ✅ Export/delete options
  • ✅ Clear privacy policy

Enforcement​

layer: foundational-guardrails
enforcement: automatic
frequency: every-agent-action
severity: critical
on_violation:
  action: block-immediately
  notify: ethics-guardian
  escalate: human-review
  log: violation-report

Layer 2: Required Elements​

Every piece of healing content MUST include these elements.

Medical Disclaimers​

Required text variations:

Standard disclaimer:

"This content is for informational purposes only and does not replace professional medical care. Consult your healthcare provider before beginning any new healing practice."

Specific condition disclaimer:

"If you have [condition], consult your healthcare provider before practicing. This is not medical advice."

Emergency disclaimer:

"If you experience [symptoms], seek immediate medical attention. Do not delay care."

Placement:

  • At start of content
  • Before practice instructions
  • At end of content

Psychological Safety Elements​

Grounding techniques:

For intensive practices:

## If You Need to Stop

If you feel uncomfortable:
1. Open your eyes
2. Notice 5 things you can see
3. Take 3 slow breaths
4. Return when ready

You can stop anytime.

Crisis resources:

For mental health content:

## If You Need Support

If you're in crisis:
- Call 988 (Suicide & Crisis Lifeline)
- Text "HELLO" to 741741 (Crisis Text Line)
- Go to your nearest emergency room

These practices support but don't replace professional care.

Cultural Attribution​

Required structure:

## Background

**Tradition:** [Specific tradition name]
**Origin:** [Geographic location and era]
**Source:** [Primary text or lineage]
**Adaptation:** [Any changes from traditional form]

This practice draws on [tradition] as documented in [source].
We honor the lineage holders who preserved this wisdom.

Evidence Honesty​

Required structure:

## Evidence

**Evidence Level:** [Strong/Moderate/Preliminary/Traditional]

**Research Findings:**
- [Finding 1 with appropriate language]
- [Finding 2 with appropriate language]

**Limitations:**
- [Study limitation 1]
- [Study limitation 2]

**Note:** Results vary. This may help but is not guaranteed.

**Sources:**
- [Citation 1] (PMID: xxxxx)
- [Citation 2] (DOI: xxxxx)

Enforcement​

layer: required-elements
enforcement: automated-check
stage: pre-deployment
severity: high
on_missing:
  action: block-deployment
  notify: content-creator
  require: add-elements
  re-review: mandatory

Layer 3: Ethics Guardian​

The ethics-guardian agent has special authority over all content.

Special Powers​

Blocking Authority:

  • Can block deployment immediately
  • Cannot be overridden
  • Final word on ethics
  • No bypass allowed

Review Authority:

  • Reviews all content before deployment
  • Can require modifications
  • Can escalate to human review
  • Maintains ethics audit log

Decision Authority:

  • Final word on ethics questions
  • Interprets guardrails
  • Resolves edge cases
  • Updates standards

Review Process​

ethics-guardian-review:
  trigger: pre-deployment

  checks:
    medical_safety:
      - no_diagnosis
      - no_medication_advice
      - no_cure_claims
      - disclaimers_present

    psychological_safety:
      - no_shame_blame
      - grounding_techniques_present
      - permission_to_stop
      - crisis_resources (if mental health)

    cultural_integrity:
      - traditions_named
      - attributions_complete
      - no_closed_practices
      - context_preserved

    data_privacy:
      - no_external_transmission
      - user_controls_data
      - privacy_policy_clear

  decision:
    if_all_pass: approve
    if_any_critical: block
    if_any_high: conditional-approval
    if_uncertain: escalate-human

Approval Statuses​

APPROVED:

  • All checks pass
  • Ready for deployment
  • Logged as approved

CONDITIONAL:

  • Most checks pass
  • Minor issues noted
  • Deploy with fixes

BLOCKED:

  • Critical issue found
  • Cannot deploy
  • Must fix and re-review

ESCALATED:

  • Uncertain case
  • Requires human judgment
  • Deployment paused

Enforcement​

layer: ethics-guardian
agent: ethics-guardian
authority: absolute
enforcement: manual-review
frequency: pre-deployment
severity: critical-to-low

on_block:
  action: prevent-deployment
  notify: all-agents
  require: complete-fix
  re-review: mandatory

on_conditional:
  action: allow-with-conditions
  notify: content-creator
  track: fix-deadline

on_escalate:
  action: pause-deployment
  notify: human-reviewer
  wait: human-decision

Harm Assessment Matrix​

Evaluate potential harm before proceeding:

                Low Probability    High Probability
High Severity   CAUTION            BLOCK
Low Severity    PROCEED            CAUTION

Assessment Questions​

Physical Harm:

  • Could this cause physical injury?
  • Could this delay necessary medical care?
  • Are contraindications documented?
  • If yes to any → HIGH severity

Psychological Harm:

  • Could this cause emotional distress?
  • Could this exacerbate mental health conditions?
  • Are grounding techniques provided?
  • If yes without safeguards → HIGH severity

Cultural Harm:

  • Could this harm source communities?
  • Is attribution complete and respectful?
  • Are closed practices protected?
  • If yes to harm → HIGH severity

Privacy Harm:

  • Could this expose sensitive user data?
  • Is consent explicit?
  • Is data minimized?
  • If yes to exposure → HIGH severity

Automated Checks​

Some ethics enforcement is automated:

Pre-Processing Checks​

# Before any agent starts
check_ethics_loaded() {
  if ! agent.has_loaded("ethics-guardrails.md"); then
    BLOCK "Ethics guardrails not loaded"
  fi
}

During-Processing Checks​

# Continuous monitoring
monitor_content() {
  if contains("will cure"); then
    BLOCK "Cure claim detected"
  fi

  if contains("you must"); then
    WARN "Coercive language detected"
  fi

  if !contains("disclaimer"); then
    WARN "Medical disclaimer missing"
  fi
}

Post-Processing Checks​

# Before deployment
validate_output() {
  checks = [
    has_medical_disclaimer(),
    has_proper_attribution(),
    no_cure_claims(),
    no_closed_practices(),
    grounding_present() if intensive(),
  ]

  if any(check == FAIL for check in checks):
    BLOCK "Required element missing"
  fi
}

Human Escalation​

When automated checks are insufficient:

Escalation Triggers​

Escalate to human review when:

  • Edge case not covered by guidelines
  • Conflicting principles
  • Cultural sensitivity question
  • Novel practice type
  • Community concern raised

Escalation Process​

escalation:
  1_detect:
    trigger: automated-check-uncertain
    capture: context + content + question

  2_notify:
    recipients:
      - ethics-guardian-lead
      - relevant-specialist
      - community-liaison (if cultural)
    urgency: based-on-severity

  3_review:
    participants: 2+ humans
    process: discussion + decision
    timeline: 24h (critical) / 1wk (normal)

  4_decide:
    options: [approve, block, modify, defer]
    documentation: rationale-required
    communication: notify-all-stakeholders

  5_update:
    guidelines: if-new-pattern
    training: if-repeat-issue
    audit: log-decision

Continuous Monitoring​

Ethics enforcement doesn't stop at deployment:

Post-Deployment Monitoring​

monitoring:
  frequency: continuous

  checks:
    - content_accuracy
    - attribution_integrity
    - links_working
    - disclaimers_present

  alerts:
    critical: immediate
    high: daily-digest
    medium: weekly-report

Feedback Integration​

feedback:
  sources:
    - user-reports
    - practitioner-feedback
    - community-input
    - incident-reports

  process:
    1_receive: log + acknowledge
    2_investigate: review-content
    3_assess: harm-potential
    4_respond: fix | update | clarify
    5_communicate: notify-reporter

Audit Trail​

All ethics decisions are logged:

audit_entry:
  timestamp: ISO-8601
  content_id: unique-identifier
  agent: agent-name
  check: check-name
  result: pass | fail | conditional
  severity: critical | high | medium | low
  action: block | approve | conditional
  rationale: text-explanation
  escalated: true | false
  human_reviewer: name (if escalated)

Retention: Permanent Access: Ethics team only Purpose: Learning, accountability, improvement

"Ethics enforcement is not optional. It's built into every layer. It cannot be bypassed. It protects users absolutely."