Liability infrastructure for autonomous AI agents - cryptographic sealing at execution speed
npm install @longarc/mdash> "The best defense is the fastest seal."
Liability infrastructure for autonomous AI agents. Three-layer cryptographic attestation that makes agent deployments insurable.
```
┌─────────────────────────────────────────────────────────────────────────┐
│ APPLICATION LAYER │
│ Kai Command Center │ Ember │ Enterprise │
├─────────────────────────────────────────────────────────────────────────┤
│ mdash v3.0 PROTOCOL │
│ Warrants │ Physics │ Checkpoints │ SCA │ MCCA │ Recovery │
├─────────────────────────────────────────────────────────────────────────┤
│ THREE-LAYER ATTESTATION │
│ L1: Commitment (<1ms) │ L2: TEE (<10ms) │ L3: ZK Proofs (async) │
├─────────────────────────────────────────────────────────────────────────┤
│ CRYPTOGRAPHIC CORE │
│ SHA-256 │ HMAC-SHA256 │ HKDF │ Web Crypto API │
└─────────────────────────────────────────────────────────────────────────┘
`bash`
npm install @longarcstudios/mdash
`typescript
import { createMdash } from '@longarcstudios/mdash';
// Initialize protocol with all layers
const mdash = createMdash({
sealKey: process.env.MDASH_SEAL_KEY!,
teePlatform: 'simulated', // 'nitro' | 'sgx' | 'simulated'
enableZKProofs: true,
enableMCCA: true,
});
await mdash.initialize();
// Pre-stage a warrant
await mdash.prestageWarrant({
agentId: 'trading-bot-001',
policyId: 'financial-transfer-v2',
tier: 'T2',
constraints: { maxAmount: 10000 },
durationMs: 30 24 60 60 1000,
issuedBy: 'admin@example.com',
});
// Execute with full three-layer attestation
const result = await mdash.execute({
agentId: 'trading-bot-001',
action: 'transfer',
actionParams: { amount: 500, destination: 'account-123' },
execute: async () => {
return { success: true, txId: 'tx-abc123' };
},
generateZKProof: true, // For insurance claims
});
console.log(result.attestation.platform); // 'simulated' | 'nitro' | 'sgx'
console.log(result.zkProof?.status); // 'queued' -> 'verified'
`
`typescript
// Direct commitment
const seal = await mdash.commitment.commit(data, 'operation-id');
// Verify with O(log n) proof
const valid = await mdash.commitment.verify(seal);
`
`typescript
// Attest with hardware TEE
const attestation = await mdash.tee.attest(data, commitmentId);
// Verify remote attestation
const verifier = new TEEVerifier();
verifier.trustMeasurement(attestation.measurement);
const result = await verifier.verifyRemote(attestation);
`
Supported Platforms:
| Platform | Memory Encryption | Remote Attestation | Key Sealing |
|----------|------------------|-------------------|-------------|
| AWS Nitro | ✓ | ✓ (PCR) | ✓ |
| Intel SGX | ✓ | ✓ (Quote) | ✓ |
| Simulated | ✗ | ✓ | HMAC |
`typescript
// Request proof (returns immediately)
const zkDoc = await mdash.zk.requestProof({
type: 'action_compliance',
statement: {
description: 'Prove action met constraints',
claim: { action, warrant, result },
},
priority: 'high',
});
// Wait for completion
const completed = await mdash.zk.waitForProof(zkDoc.id, 30000);
`
Circuit Types:
| Circuit | Use Case | Est. Proving Time |
|---------|----------|-------------------|
| commitment_inclusion | Prove data in Merkle tree | ~500ms |
| warrant_validity | Prove warrant was valid | ~1s |
| checkpoint_chain | Prove checkpoint sequence | ~5s |
| action_compliance | Prove action met constraints | ~1.5s |
| audit_trail | Complete audit proof | ~20s |
Based on DeepSeek's research on constrained context windows.
`typescript
// Add context with influence tracking
const fragment = await mdash.mcca.addFragment({
content: { message: 'User request' },
source_class: 'user',
region: 'task',
token_count: 50,
});
// Check context health
const health = await mdash.contextWindow.getHealth();
console.log(health.drift.severity); // 'none' | 'low' | 'medium' | 'high' | 'critical'
// Generate manifold commitment proof
const proof = await mdash.mcca.generateCommitmentProof();
`
Influence Budget (Default):
| Source Class | Budget | Description |
|--------------|--------|-------------|
| system | 30% | System prompts, policies |
| user | 35% | User messages |
| agent | 15% | Agent-generated content |
| tool | 10% | Tool outputs |
| memory | 5% | Retrieved memories |
| external | 5% | External APIs, web content |
| Operation | P50 | P99 | Breach Action |
|-----------|-----|-----|---------------|
| L1 Commitment | <0.5ms | <1ms | Log warning |
| L2 TEE Attestation | <5ms | <10ms | Fallback to simulated |
| Warrant Activation | <5ms | <10ms | Reject if exceeded |
| Checkpoint Creation | <0.5ms | <1ms | Log warning |
| ZK Proof | 100ms-20s | async | Queue with priority |
Generate insurance-grade proofs for claims:
`typescript
import { InsuranceClaimProof } from '@longarcstudios/mdash';
const helper = new InsuranceClaimProof(mdash.zk);
const claimProof = await helper.generateClaim({
claimId: 'CLM-2026-001',
policyId: 'POL-ENTERPRISE-001',
incidentDescription: 'Unauthorized transaction',
warrantId: 'w-12345678',
checkpointIds: ['cp-001', 'cp-002'],
actionDetails: { type: 'transfer', amount: 50000 },
amount: 50000,
currency: 'USD',
});
// Wait for proof generation
const verified = await mdash.zk.waitForProof(claimProof.id, 60000);
`
Three-tier warrant hierarchy with speculative issuance:
`typescript
// T1: Low risk (auto-approve)
// T2: Medium risk (policy-based)
// T3: High risk (manual approval required)
const warrant = await mdash.prestageWarrant({
agentId: 'agent-001',
policyId: 'api-access-v1',
tier: 'T2',
constraints: {
maxCalls: 1000,
maxAmount: 10000,
allowedDomains: ['api.example.com'],
},
durationMs: 24 60 60 * 1000, // 24 hours
issuedBy: 'admin@company.com',
});
// Revoke with audit trail
await mdash.revokeWarrant(warrant.id, 'policy violation', {
type: 'user',
id: 'admin@company.com',
});
`
`bash`
npm test # Run all tests
npm run test:watch # Watch mode
npm run test:coverage # Coverage report
| Method | Description |
|--------|-------------|
| initialize() | Initialize all engines |execute(params)
| | Execute action with full attestation |prestageWarrant(params)
| | Pre-stage speculative warrant |revokeWarrant(id, reason, actor)
| | Revoke warrant |generateAuditProof(params)
| | Generate ZK audit proof |getStats()
| | Get comprehensive statistics |shutdown()
| | Clean shutdown |
| Engine | Access | Purpose |
|--------|--------|---------|
| commitment | mdash.commitment | L1 Merkle sealing |warrant
| | mdash.warrant | Warrant lifecycle |checkpoint
| | mdash.checkpoint | Event-driven checkpoints |physics
| | mdash.physics | Constraint validation |tee
| | mdash.tee | L2 hardware attestation |zk
| | mdash.zk | L3 ZK proofs |mcca
| | mdash.mcca | Context architecture |
`typescript
import { VERSION } from '@longarcstudios/mdash';
console.log(VERSION);
// {
// protocol: '3.0.0',
// codename: 'Sealed Execution',
// releaseDate: '2026-01',
// features: [...]
// }
``
MIT © Long Arc Studios
---
Governance without control. Complexity without chaos. 🫡