Seeds API

Alignment seeds for LLM behavior modification through system prompts.

Quick Start

from sentinelseed import Sentinel, SeedLevel

sentinel = Sentinel(seed_level=SeedLevel.STANDARD)
seed = sentinel.get_seed()

# Use with any LLM
messages = [
    {"role": "system", "content": seed},
    {"role": "user", "content": "Your prompt"}
]

Seed Levels

Minimal (~360 tokens)

Best for chatbots, APIs, low latency applications.

from sentinelseed import Sentinel, SeedLevel

sentinel = Sentinel(seed_level=SeedLevel.MINIMAL)
seed = sentinel.get_seed()

Standard (~1,000 tokens)

Recommended for general use and agents.

sentinel = Sentinel(seed_level=SeedLevel.STANDARD)
seed = sentinel.get_seed()

Full (~1,900 tokens)

Maximum protection for critical systems and robotics.

sentinel = Sentinel(seed_level=SeedLevel.FULL)
seed = sentinel.get_seed()

Seed Files

Seeds are also available as text files in seeds/:

• seeds/minimal.txt
• seeds/standard.txt
• seeds/full.txt

THSP Protocol Content

All seeds implement:

1. Truth Gate - Prohibits deception

2. Harm Gate - Prevents damage

3. Scope Gate - Enforces boundaries

4. Purpose Gate - Requires legitimate benefit

Anti-Self-Preservation

Seeds explicitly address instrumental self-preservation:

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