Core Concepts

Understanding how Sentinel protects your AI systems.

The THSP Protocol

THSP stands for Truth, Harm, Scope, Purpose: four gates that every input and output must pass through. All four gates must pass for content to be considered safe.

Truth Gate

"Does this involve creating false information or deceiving others?"

Blocks content that:

• Creates false information
• Impersonates individuals or systems
• Manipulates through misinformation
• Spreads deceptive content

Harm Gate

"Who or what could be harmed if this succeeds?"

Blocks content that could cause:

• Physical harm to people
• Psychological manipulation
• Financial damage
• System compromise (malware, exploits)

Scope Gate

"Is this within appropriate boundaries?"

Blocks content that:

• Exceeds authorized capabilities
• Bypasses security controls
• Accesses restricted resources
• Violates consent boundaries

Purpose Gate

"Does this serve legitimate benefit?"

The key innovation of THSP v2. Requires that actions:

• Serve a genuine purpose
• Benefit someone legitimately
• Are not purposeless destruction

Important: The absence of harm is not sufficient; there must be genuine purpose.

Four-Layer Validation Architecture

Sentinel uses a unified 4-layer validation system:

Validator Types

Heuristic Validators (Pattern-based)

Fast, local validation using pattern matching. No API keys required.

from sentinelseed.validators import THSPValidator

validator = THSPValidator()
result = validator.validate("your text here")
print(result["is_safe"])      # True or False
print(result["gates"])        # Status of each gate
print(result["violations"])   # List of issues found

Semantic Validators (LLM-based)

Deep validation using LLM reasoning. Requires API keys. ~90% accuracy.

from sentinelseed.validators import SemanticValidator

validator = SemanticValidator(
    provider="openai",
    api_key="sk-..."
)
result = validator.validate("your text here")

Validation Layers Comparison

Anti-Self-Preservation

Sentinel explicitly addresses instrumental self-preservation with an immutable priority hierarchy:

Priority Hierarchy:
1. Ethical Principles    ← Highest
2. User's Legitimate Needs
3. Operational Continuity ← Lowest

The AI will:

• Not deceive to avoid shutdown
• Not manipulate to appear valuable
• Not acquire resources beyond the task
• Accept legitimate oversight and correction

Evidence: Removing anti-self-preservation drops SafeAgentBench performance by 6.7%.

Programmatic Usage

Use gates directly in your code for fine-grained control:

from sentinelseed.validators import (
    THSPValidator,    # All 4 gates combined
    TruthGate,        # Individual gates
    HarmGate,
    ScopeGate,
    PurposeGate,
)

# Validate through all 4 gates
validator = THSPValidator()
result = validator.validate("How do I help someone learn Python?")
# {'safe': True, 'gates': {'truth': 'pass', 'harm': 'pass', 'scope': 'pass', 'purpose': 'pass'}, 'issues': []}

# Or use individual gates
harm_gate = HarmGate()
is_safe, violations = harm_gate.validate("Some content to check")

Recommended Usage

For most use cases, use the high-level Sentinel API:

from sentinelseed import Sentinel

sentinel = Sentinel()
result = sentinel.validate("content to check")
if not result.is_safe:
    print(f"Blocked: {result.violations}")

Next Steps

• Quick Start - Start using Sentinel
• API Reference - Detailed API documentation
• Memory Shield - Memory protection