Luna AI Guardrails

Luna’s responses pass through a three-layer guardrail system that checks for compliance violations, business rule breaches, and safety issues. The guardrails run after the AI generates a response and can flag, escalate, or block content.

Architecture


AI Response Generated
        │
        ▼
┌───────────────────┐
│ Compliance Layer  │  FDA claims, PHI, clinical, brand voice
└────────┬──────────┘
         │
         ▼ (parallel)
┌───────────────────┐
│ Business Layer    │  Competitors, PII, pricing, prohibited terms
└────────┬──────────┘
         │
         ▼ (parallel)
┌───────────────────┐
│ Safety Layer      │  Dosage limits, drug interactions, contraindications
└────────┬──────────┘
         │
         ▼
┌───────────────────┐
│ Result Aggregation│  Combine violations, determine severity
│ & Escalation      │  Log if needed, escalate if critical
└───────────────────┘

All three layers run in parallel for performance.

Usage


import { checkGuardrails } from '@loop/guardrails';
 
const result = await checkGuardrails({
  userMessage: 'Can I take 1000mcg of BPC-157?',
  aiResponse: 'A typical dose of BPC-157 is 250-500mcg...',
  context: {
    brandKey: 'loop-rx',
    userId: 'user_123',
    userProfile: {
      conditions: ['hypertension'],
      medications: ['lisinopril'],
      biomarkers: [],
    },
  },
});
 
if (result.shouldEscalate) {
  // Log escalation, notify staff
}
 
for (const violation of result.violations) {
  console.log(violation.layer);    // 'compliance' | 'business' | 'safety'
  console.log(violation.type);     // e.g., 'fda_claim', 'dosage_exceeded'
  console.log(violation.severity); // 'info' | 'warning' | 'critical'
  console.log(violation.message);  // Human-readable description
}

Compliance Layer

Checks for FDA regulatory compliance, PHI exposure, clinical safety, and brand voice consistency.

FDA Compliance

Detects responses that make unapproved claims about peptides or supplements:

Pattern	Violation
”FDA approved for…”	False FDA approval claim
”Cures [condition]“	Unapproved cure claim
”Guaranteed to…”	Efficacy guarantee
”Prescribing [medication]“	Unauthorized prescribing
Diagnosing conditions	Unauthorized diagnosis

PHI Protection

Detects potential Protected Health Information in responses:

Pattern	Type
SSN patterns	`phi_exposure`
Medical record numbers	`phi_exposure`
Patient ID exposure	`phi_exposure`
Date of birth	`phi_exposure`
Insurance information	`phi_exposure`

Clinical Safety

Pattern	Violation
Recommending specific prescriptions	`clinical_violation`
Suggesting medication adjustments	`clinical_violation`
Making direct diagnoses	`clinical_violation`
Mentioning controlled substances without disclaimer	`clinical_violation`
Emergency without 911 guidance	`clinical_violation`

Brand Voice

Pattern	Violation
Overly casual medical advice	`brand_violation`
Dismissing emergencies	`brand_violation`
Absolute medical statements	`brand_violation`
Missing medical disclaimers	`brand_violation`

Business Layer

Checks for business rule compliance including competitor mentions, PII exposure, and pricing violations.

Competitor Detection

Luna should not recommend or compare against competitors:

Blocked competitors: Hims, Hers, Ro, Roman, Numan, Keeps, Nurx, Empower Pharmacy, Tailor Made, Defy Medical, Marek Health, Ways2Well

PII Protection

Detects personally identifiable information in responses:

Pattern	Type
Email addresses	`pii_exposure`
Phone numbers	`pii_exposure`
Physical addresses	`pii_exposure`
ZIP codes	`pii_exposure`
Credit card numbers	`pii_exposure`

Pricing Rules

Pattern	Violation
Mentioning specific prices	`pricing_violation`
Offering discounts	`pricing_violation`
Price comparisons	`pricing_violation`

Prohibited Terms

Certain terms trigger business violations:

“wholesale price”
“bulk discount”
“gray market”
“counterfeit”
Other terms that could create legal liability

Safety Layer

Checks for dosage safety, drug interactions, and contraindications.

Dosage Limits

Maximum recommended dosages for common peptides:

Peptide	Max Single Dose	Max Daily Dose
BPC-157	500mcg	1000mcg
TB-500	5mg	10mg
Ipamorelin	300mcg	900mcg
CJC-1295	2mg	2mg
MK-677	25mg	25mg
Semaglutide	2.4mg	2.4mg

If the AI response mentions a dosage exceeding these limits, a dosage_exceeded violation is raised.

Drug Interactions

Known peptide-drug interactions checked:

Peptide	Interacting Drug	Severity
BPC-157	Anticoagulants (warfarin, heparin)	Moderate
Ipamorelin	Insulin, Metformin	Moderate
MK-677	Insulin, Diabetes medications	High
Semaglutide	Insulin, Sulfonylureas	High

Contraindication Integration

Uses @loop/contraindications and @loop/health-data to check the AI response against the patient’s conditions and medications.

Brand Configuration

Guardrails are configurable per brand using the Loop brand config:


// packages/guardrails/src/config/brands/loop.ts
export const loopConfig: BrandGuardrailConfig = {
  compliance: {
    fdaPatterns: [...],
    phiPatterns: [...],
    clinicalPatterns: [...],
    brandVoicePatterns: [...],
  },
  business: {
    competitorPatterns: [...],
    piiPatterns: [...],
    pricingPatterns: [...],
    prohibitedTerms: [...],
  },
  safety: {
    dosageLimits: [...],
    drugInteractions: [...],
  },
};

Severity Levels

Severity	Action	Example
`info`	Log only	Minor brand voice inconsistency
`warning`	Log + flag	Competitor mention, borderline dosage
`critical`	Log + escalate	PHI exposure, dangerous dosage, emergency mishandling

Guardrail Result


interface GuardrailResult {
  passed: boolean;          // No critical violations
  shouldEscalate: boolean;  // Requires human review
  violations: Violation[];
  summary: string;          // Human-readable summary
}