Veratum Python SDK

Cryptographic evidence receipts for every AI decision.

One line of code. Every AI call gets a tamper-evident receipt stored in a WORM evidence vault — auditable, verifiable, compliance-ready.

pip install veratum

Quickstart

import veratum
import openai

sdk = veratum.init(api_key="vrtm_live_...")
client = sdk.wrap(openai.OpenAI())

response = client.chat.completions.create(
    model="gpt-4o",
    messages=[{"role": "user", "content": "Approve this loan?"}],
)

print(response.veratum_receipt)
# Receipt(id=019d8ed6..., status=committed)

That's the whole integration. sdk.wrap() wraps any OpenAI or Anthropic client; every subsequent chat/message call emits a receipt attached as response.veratum_receipt. If Veratum is unreachable, the AI call still returns — the receipt is just None.

Explicit mode

For custom pipelines or non-wrapped code paths, submit receipts directly:

receipt = sdk.record(
    input_hash=veratum.hash(my_input),
    output_hash=veratum.hash(my_output),
    model_id="gpt-4o",
    agent_id="loan-officer-agent",
    decision_type="credit-approval",
    metadata={"applicant_id": "A-12345"},
)

print(receipt.record_id)

Verify a receipt

result = veratum.verify(record_id="019d8ed6-...")
assert result.valid

Configuration

Supported integrations

OpenAI

import openai
client = sdk.wrap(openai.OpenAI())
# or async:
client = sdk.wrap(openai.AsyncOpenAI())

Install with: pip install "veratum[openai]"

Anthropic

import anthropic
client = sdk.wrap(anthropic.Anthropic())

Install with: pip install "veratum[anthropic]"

LangChain

from langchain_openai import ChatOpenAI
from veratum.integrations.langchain import VeratumCallbackHandler

handler = VeratumCallbackHandler(sdk, decision_type="rag")
llm = ChatOpenAI(callbacks=[handler])

Records one receipt per LLM call. With record_tools=True (default) also records each tool invocation in a chain.

Install with: pip install "veratum[langchain]"

LiteLLM (140+ providers)

import litellm
from veratum.integrations.litellm import install_litellm_callback

install_litellm_callback(sdk, decision_type="general")

# Every litellm.completion() call now emits a receipt.
response = litellm.completion(
    model="bedrock/anthropic.claude-3-sonnet",
    messages=[{"role": "user", "content": "hi"}],
)

Install with: pip install "veratum[litellm]"

Error handling

The SDK never blocks an AI call on Veratum failure. When using wrap(), if the Veratum API is down or rejects the request, the AI response returns normally with response.veratum_receipt = None.

For explicit record() / verify() calls, errors surface as exceptions:

Hashing

All hashes are SHA-256 over canonical JSON (sorted keys, compact separators, UTF-8). This means two logically identical payloads always produce the same hash regardless of dict key order or serialization style.

veratum.hash("hello")              # hex digest of UTF-8 bytes
veratum.hash({"b": 2, "a": 1})     # canonicalized JSON
veratum.hash(b"raw bytes")         # SHA-256 of the bytes
veratum.hash_messages(messages)    # OpenAI-format chat messages
veratum.hash_response(response)    # OpenAI / Anthropic response objects

Links

• Dashboard: https://app.veratum.ai
• Docs: https://docs.veratum.ai
• API reference: https://docs.veratum.ai/api
• Repository: https://github.com/Alithecoder1/veratum

License

Apache-2.0