From zero to first signed decision.

Four lines of Python, one decision record. This is what onboarding to Veridra actually looks like.

Before you start

• A Veridra sandbox API key (request through contact during pre-GA; self-serve sign-up arrives with GA).
• A signing key — for sandbox, we provision a managed key; for production, BYOK through your KMS.
• One model call you want to wrap — any LLM provider, any custom model, any decision you want proved.

Step 1 — Install the SDK

Pick the language you're working in. All four SDKs expose the same conceptual interface.

pip install veridra

npm install @veridra/sdk

go get github.com/veridra/sdk-go@latest

<dependency>
  <groupId>io.veridra</groupId>
  <artifactId>veridra-sdk</artifactId>
  <version>0.1.0</version>
</dependency>

Step 2 — Wrap your first decision

The pattern is always the same: construct a client, call attest() around your model invocation, receive a signed decision record back. Here is the Python example:

from veridra import Client

veridra = Client(api_key="vk_sandbox_...")

# Your existing model call
def underwrite(application):
    result = your_model.predict(application)
    return result

# Wrap it
application = {"income": 85000, "credit_score": 720, "state": "NY"}

with veridra.attest(
    system="loan-underwriter-v3.2",
    risk_tier="high",              # EU AI Act Annex III
    jurisdiction="US-NY",
    inputs=application,
) as decision:
    result = underwrite(application)
    decision.record(
        outcome=result["approved"],
        confidence=result["confidence"],
        policy_checks=["fair-lending", "adverse-action"],
        human_review="not_required",
    )

print(decision.id)            # dec_9a3f21b4
print(decision.signed)        # True
print(decision.log_entry)     # 18942017

Step 3 — Receive the signed record

The record your SDK produces is the canonical Veridra decision object. Every field is signed together; tampering with any of them breaks the signature.

{
  "decision_id": "dec_9a3f21b4",
  "system": "loan-underwriter-v3.2",
  "risk_tier": "high",
  "outcome": "approve",
  "confidence": 0.874,
  "inputs_hash": "sha256:7f2a...c091",
  "model_hash": "sha256:e4d1...b8a3",
  "policy_checks": ["fair-lending", "adverse-action"],
  "human_review": "not_required",
  "timestamp": "2026-04-20T14:22:07.441Z",
  "jurisdiction": "US-NY",
  "signature": "ed25519:5d3b...7a02",
  "log_entry": 18942017
}

Step 4 — Verify it with the open-source CLI

The verification CLI is open-source (Apache 2.0), runs locally, and needs no network connection to Veridra. This is the property that lets customers, regulators, and auditors independently confirm the evidence.

# Install
brew install veridra-verify           # macOS
apt install veridra-verify            # Debian / Ubuntu

# Verify a single record
veridra-verify record dec_9a3f21b4.json
# ✓ signature valid · ed25519
# ✓ canonical form matches
# ✓ transparency log inclusion proof verified
# ✓ record is internally consistent

# Verify an entire evidence pack
veridra-verify pack evidence-2026-04.zip

Where to go next

• Wire through your KMS: switch from the sandbox managed key to BYOK in AWS KMS, Azure Key Vault, GCP KMS, or Vault. See SDK configuration.
• Add policy checks: declare your governance policies in Rego and have them enforced at the signing gateway.
• Add continuous monitoring: enable Watch module to get drift detection and scheduled evaluation on the same decision stream.
• Build your first evidence pack: generate a regulator-ready pack with veridra pack once you have decision volume.
• Integrate with your CI: sign your model cards as part of the deployment pipeline using the same SDK.

Provider-specific wrappers

For common LLM providers, we ship drop-in wrappers that handle the integration boilerplate:

• veridra.integrations.openai — wraps OpenAI Python SDK with attest
• veridra.integrations.anthropic — wraps Anthropic Python SDK with attest
• veridra.integrations.bedrock — AWS Bedrock InvokeModel calls
• veridra.integrations.vertex — Google Vertex AI prediction requests