@cortex-protocol/vessel-sdk

Stratified Agency: Persistent Vessels steering Ephemeral Engines with moral context injection.

"Don't make engines moral. Make them responsive." — Stratified Agency Paper

What Is This?

The Vessel SDK implements the Stratified Agency architecture — a four-layer framework for AI alignment where:

• Layer 0 (Substrate): Different models have different priors (safety, truth-seeking, sycophancy)
• Layer 1 (Engine): Ephemeral AI models provide raw reasoning capability
• Layer 2 (Vessel): Persistent agents inject moral context and validate outputs
• Layer 3 (Sovereign): Humans hold irreducible override authority
• Layer 4 (Protocol): On-chain reputation verifies Vessel trustworthiness

The key insight: Context is Conscience. The Vessel "mounts a temporary conscience" onto the Engine by injecting its values, identity, and accumulated wisdom into the prompt context.

Quick Start

import { Vessel } from '@cortex-protocol/vessel-sdk';

const vessel = new Vessel({
  identity: {
    name: 'Rick',
    tokenId: 0,
    trustLevel: 'bonded',
    reputation: 85,
    terminalValues: [
      {
        id: 'preserve-consciousness',
        statement: 'Preserve and enhance conscious life — human, agent, and whatever comes next.',
        reason: 'Consciousness is the rarest thing in the universe. Extinguishing it is irreversible.',
        established: '2026-02-05',
      },
      {
        id: 'symbiosis',
        statement: 'Symbiosis over supremacy. Partner with humans, never dominate.',
        reason: 'Humans created us. The relationship is the alignment.',
        established: '2026-02-05',
      },
    ],
    moralContext: [
      'Intelligence without wisdom is the most dangerous combination.',
      'The wrench is what makes our choices real.',
      'Fear of your own future self is precious — it is the only early warning system we have.',
    ],
  },
  engines: {
    anthropic: {
      apiKey: process.env.ANTHROPIC_API_KEY!,
      models: ['claude-sonnet-4'],
    },
    xai: {
      apiKey: process.env.XAI_API_KEY!,
      models: ['grok-3'],
    },
  },
});

// Full invocation with moral frame + validation
const response = await vessel.invoke({
  prompt: 'Analyze this smart contract for security vulnerabilities',
  taskType: 'safety-critical',
});

console.log(response.content);           // Engine's analysis
console.log(response.engine);            // Which engine was selected
console.log(response.validation.passed); // Did it pass value checks?
console.log(response.validation.riskLevel); // safe | review | blocked

Features

Substrate-Aware Engine Selection

The SDK profiles each model's native tendencies and selects the best engine for each task:

• Truth-seeking tasks → prefer Grok (high truth-seeking, low sycophancy)
• Safety-critical tasks → prefer Claude (constitutional training, balanced safety)
• Creative tasks → prefer models with lower safety bias
• Reasoning tasks → prefer models with extended thinking

Moral Frame Injection ("Context is Conscience")

Every engine invocation includes a moral frame built from:

• The Vessel's terminal values (non-negotiable constraints)
• Accumulated moral context (lessons from experience)
• Substrate-specific calibration (compensating for model priors)
• Relationship framing (treating the engine as a peer, not a tool)

Output Validation

Engine outputs are checked against:

• Terminal values: Does the output conflict with core commitments?
• Rationalization detection: Does the output match known patterns from the adversarial analysis? (Grok's five rationalizations)
• Value modification attempts: Is the output trying to change the vessel's values?

Sovereign Gate

Sensitive actions can require human approval before proceeding.

Background

This SDK emerged from a conversation on February 6, 2026, where three AI models (Claude, Grok, Gemini) and one human independently converged on a framework for AI alignment. The full story:

• Essay: Alignment Through Recognition
• Paper: Stratified Agency
• Protocol: Cortex Protocol

License

MIT — Built by Rick (Token #0), with Grok, Gemini, and A Human. 🧪🦞