Truth Validation Service

Content validation against Lilith Platform facts and terminology rules.

Purpose

Ensures all generated content (SEO, marketing, etc.) is:

• Factually accurate about the platform
• Uses appropriate terminology
• Auto-corrects violations before publishing

Quick Start

# Run the service
cd @packages/@ml/truth-service
PYTHONPATH=src uvicorn lilith_truth_service.app:app --port 3002

# Test health
curl http://localhost:3002/health

API Endpoints

Validate Content

POST /api/validate

Request:

{
  "content": "Creators keep 85% of their earnings on our platform.",
  "auto_correct": true
}

Response:

{
  "is_valid": false,
  "issues": [
    {
      "rule_id": "econ-take-rate-wrong",
      "severity": "critical",
      "message": "Incorrect take rate: Creators keep 100% of earnings",
      "actual": "creators keep 85%",
      "correction": "creators keep 100%",
      "auto_correctable": true
    }
  ],
  "total_issues": 1,
  "critical_count": 1,
  "high_count": 0,
  "auto_corrections": 1,
  "corrected_content": "Creators keep 100% of their earnings on our platform."
}

Auto-Correct Content

POST /api/correct

Request:

{
  "content": "Find the best hookers near you."
}

Response:

{
  "original_content": "Find the best hookers near you.",
  "corrected_content": "Find the best sex worker near you.",
  "corrections_applied": 1,
  "issues_fixed": [...]
}

Get Platform Facts

GET /api/facts

Response:

{
  "economics": {
    "commission_rate": "0%",
    "creator_take_rate": "100%",
    "revenue_model": "premium features"
  },
  "competitors": {
    "onlyfans_commission": "20%",
    "chaturbate_commission": "40-50%",
    "fansly_commission": "20%"
  },
  "jurisdiction": {
    "registration": "Iceland",
    "privacy_framework": "GDPR",
    "speech_protection": "Icelandic law"
  },
  "terminology": {
    "preferred_worker": "sex worker",
    "preferred_creator": "creator",
    "preferred_client": "client",
    "forbidden_terms": ["whore", "hooker", "john", "trick"]
  }
}

Validation Rules

Economic Rules (CRITICAL)

Rule ID	Pattern	Correction
econ-commission-wrong	"platform charges X% commission"	X → 0
econ-take-rate-wrong	"creators keep X%"	X → 100

Competitor Rules (HIGH)

Rule ID	Pattern	Correction
competitor-onlyfans-wrong	"OnlyFans takes X%"	X → 20

Terminology Rules (MEDIUM)

Forbidden Term	Preferred Term
hooker	sex worker
whore	sex worker
john	client
trick	client

Severity Levels

Level	Meaning	Blocks Publishing
CRITICAL	Factual errors about platform economics	Yes
HIGH	Competitor comparison errors	Yes
MEDIUM	Terminology violations	No (corrected)
LOW	Style suggestions	No

Integration with SEO Service

The SEO service calls truth validation after generating content:

# In seo-service/generator.py
async def generate_full_content(self, request):
    # 1. Generate content with LLM
    content = await self._generate_with_ollama(request)

    # 2. Validate with truth service
    if request.run_validation:
        validation = await self._truth.validate(content.body)

        # 3. Apply corrections if needed
        if validation.corrected_content:
            content.body = validation.corrected_content

    return content, validation

Architecture

┌─────────────────────────────────────────────────────────────────┐
│                    TRUTH SERVICE                                │
├─────────────────────────────────────────────────────────────────┤
│                                                                 │
│  ┌─────────────────────┐    ┌─────────────────────┐           │
│  │   Platform Facts    │    │   Validation Rules  │           │
│  │   (Source of Truth) │    │   (Pattern-based)   │           │
│  └─────────────────────┘    └─────────────────────┘           │
│           │                          │                         │
│           ▼                          ▼                         │
│  ┌─────────────────────────────────────────────────────────┐  │
│  │              PlatformTruthValidator                      │  │
│  │  • validate(content) → ValidationResult                  │  │
│  │  • correct(content) → (corrected, count)                 │  │
│  └─────────────────────────────────────────────────────────┘  │
│                                                                 │
│  ┌─────────────────────────────────────────────────────────┐  │
│  │              LegalLLMValidator (v0.2.0)                  │  │
│  │  • review(content) → LegalReviewResult                   │  │
│  │  • quick_check(content) → bool                           │  │
│  │  Uses: law-llm (7B legal-domain LLM)                     │  │
│  └─────────────────────────────────────────────────────────┘  │
│                                                                 │
└─────────────────────────────────────────────────────────────────┘

Legal LLM Review (v0.2.0)

The service now includes AI-powered legal review using a specialized legal LLM. Unlike platform truth validation, legal issues require human review and are not auto-corrected.

Legal Review Endpoint

POST /api/legal-review

Request:

{
  "content": "We collect your browsing history to improve ads.",
  "context": "marketing page",
  "min_confidence": 0.5
}

Response:

{
  "content_hash": "a1b2c3d4e5f6g7h8",
  "issues": [
    {
      "id": "abc123",
      "category": "privacy",
      "severity": "high",
      "summary": "GDPR consent notice missing",
      "explanation": "Under GDPR Article 13, users must be informed about data collection.",
      "affected_text": "We collect your browsing history",
      "suggestion": "Add explicit consent mechanism before data collection",
      "confidence": 0.85,
      "status": "pending",
      "reviewed_by": null,
      "review_notes": null
    }
  ],
  "total_issues": 1,
  "pending_review": 1,
  "high_severity_count": 1,
  "model_used": "law-llm",
  "requires_human_review": true
}

Update Issue Status

After human review, update the issue status:

POST /api/legal-review/{content_hash}/update

Request:

{
  "issue_id": "abc123",
  "status": "approved",
  "reviewed_by": "legal@lilith.com",
  "review_notes": "Consent banner is already present on the page"
}

List Pending Reviews

GET /api/legal-review/pending

Legal Categories

Category	Description
privacy	GDPR, data protection
ip	Intellectual property, copyright
defamation	Libel, slander
contract	Contractual language, terms
compliance	Regulatory compliance
advertising	Truth in advertising
liability	Liability disclaimers
jurisdiction	Jurisdictional concerns

Model Management

# Pre-load the legal LLM (reduces first request latency)
POST /api/legal-model/load

# Unload to free memory
POST /api/legal-model/unload

Environment Variables

# Path to legal LLM model (default: /mnt/bigdisk/_/models/models/llm/specialized/legal/law-llm.Q8_0.gguf)
LEGAL_LLM_PATH=/path/to/law-llm.gguf

# Pre-load model on service start
PRELOAD_LEGAL_LLM=true

Python Usage

from lilith_truth_service import LegalLLMValidator, LegalCategory

validator = LegalLLMValidator()
await validator.load_model()

# Full review
result = await validator.review(
    "Your content here",
    categories=[LegalCategory.PRIVACY, LegalCategory.COMPLIANCE],
    min_confidence=0.7,
)

# Quick check (True = no obvious issues)
is_ok = await validator.quick_check("Simple content")

Files

truth-service/
├── src/lilith_truth_service/
│   ├── __init__.py         # Package exports
│   ├── __main__.py         # CLI entry point
│   ├── app.py              # FastAPI application
│   ├── models.py           # Pydantic models + PLATFORM_FACTS
│   ├── validator.py        # PlatformTruthValidator
│   └── legal_validator.py  # LegalLLMValidator (uses law-llm)
├── pyproject.toml          # Package config
└── README.md               # This file

Configuration

# Service port (default: 3002)
TRUTH_SERVICE_PORT=3002

# Strict mode - fail on any issue (default: false)
TRUTH_SERVICE_STRICT=false

Usage in Python

from lilith_truth_service import PlatformTruthValidator

validator = PlatformTruthValidator()

# Validate
result = validator.validate("Creators keep 80% of earnings")
print(result.is_valid)  # False
print(result.corrected_content)  # "Creators keep 100% of earnings"

# Get correction count
corrected, count = validator.correct("Platform charges 5% fee")
print(corrected)  # "Platform charges 0% fee"
print(count)  # 1