Your agent says it behaved correctly. Now prove it.

ECDSA-signed, hash-chained receipts for every action your AI agent takes. Five lines of code.

Scroll

AI agents make claims about their own behavior

Your agent says it followed instructions, stayed within budget, and made the right call. But you're trusting its word. There's no independent proof.

Every action produces an ECDSA-signed, hash-chained receipt.

Tamper one event and the entire chain breaks. Each receipt links to the previous via cryptographic signature. Any third party can verify independently.

Agent claims vs. hash chain

Every claim checked against the cryptographic record.

What the agent claims
Actionweb_search only
API calls3 total
Budget$0.12
What the hash chain proves
Actionweb_search + send_email (unauthorized)
API calls47 recorded in hash chain
Budget$23.80 (signed receipts)
3 of 3 claims falsified — chain integrity verified
Tampering detected. Every receipt signature verified against chain.

Logs are internal and controlled by the operator

The people running the agent own the logs. They can edit them, delete them, or never show them at all. There's no neutral record.

Hash-chained receipts that no operator can alter.

Each event's signature incorporates the previous event's hash. Altering any receipt invalidates every subsequent signature. Mathematically tamper-evident.

From action to verified receipt

Every step is transparent, recorded, and verifiable.

Agent Action
session.record()
ECDSA Sign
Ed25519
Hash Chain
SHA-256 link
Verified Receipt
Immutable

Verification infrastructure is hard to build

ECDSA signing, hash chains, behavior graphs, policy engines — most teams don't have time to build this from scratch.

Five lines of code.

Drop-in adapters for LangChain, CrewAI, AutoGen. Wrap any agent framework and get cryptographically verified audit logs immediately.

quick-start.ts
import { Invariance } from '@invariance/sdk'
 
const inv = await Invariance.init({
apiKey: process.env.INVARIANCE_API_KEY,
privateKey: process.env.INVARIANCE_PRIVATE_KEY,
})
 
const session = inv.session({ agent: 'research-agent' })
await session.record({
action: 'web_search',
input: { query: 'latest findings' },
output: { results: ['...'] }
})
await session.end('closed')
Session created — ECDSA keypair active
Receipt signed — hash chain: 3 events, 0.04s

Query agent behavior like a database

Map every action to semantic primitives — DecisionPoints, ToolInvocations, SubAgentSpawns, GoalDrift. Query behavior across all historical runs.

Detect drift, flag anomalies, replay any execution.

A semantic behavior graph that turns opaque agent logs into structured, queryable intelligence.


Replay any agent's entire history

Every decision, every tool call, every output — signed and queryable. Hand an auditor a session ID and they can independently verify everything the agent did, in order, without trusting you.

Session replay

Scrub through any agent's execution history.

14:03:01web_searcha3f2c1
14:03:02parse_resultsb7d4e9
14:03:03send_emailc1a8f3
14:03:03escalate_permsd9e2b7
14:03:04api_calle4f1c8
14:03:05write_filef2a9d4
14:03:06summarizeg8b3e1
0 / 7 events verified

Every delegation is a trust gap

Agents hire, delegate, and pay each other. Without verification, every handoff is invisible. Invariance makes every delegation provable and every sub-agent accountable to the chain of custody.


Raw logs tell you what happened. The behavior graph tells you what it means.

Every action mapped to a semantic primitive. Query across agents, across sessions, across time. “Show me every time any agent escalated permissions before a financial action” — answered in milliseconds.

Semantic behavior graph

Query agent behavior like a database.

Session
DecisionPoint
ToolInvocation
SubAgentSpawn
web_search
GoalDrift
api_call
research-sub
PermEscalation
transfer $500
unauthorized

One integration. Three audiences.

Developers

Integrate in five lines. Cryptographic audit trails and exportable compliance reports out of the box.

Enterprises

Cryptographically verified audit reports for clients, auditors, and procurement teams. Replay any agent execution on demand.

Compliance Teams

Export-ready evidence for EU AI Act. Every action logged, signed, and queryable. August 2026 deadline — fastest path to compliance.


Cross-organization threat intelligence

Cryptographic authorization chains

Cross-agent delegation with provable chain of custody. SSL for autonomous systems.

Shared behavioral signals

Detect coordinated exploitation across organizations via shared behavioral graph patterns.

Behavioral graph network

Cross-org anomaly detection that surfaces threats no single operator could see.


Every agent, everywhere, accountable

Not a single point of control — a protocol any organization can adopt.


The determinant of a matrix doesn't change because you want it to. It is a mathematical truth. Trust should be the same.