Tamper-evident audit logs for AI systems. Hash-chained, signed, customer-owned. TypeScript + Python, one chain.

---

EU AI Act Article 12 kicks in 2 Aug 2026. Every high-risk AI system (HR, credit, insurance, healthcare, education) has to keep reconstructable, tamper-evident logs for 6+ months. Observability tools (Langfuse, Datadog LLM, LangSmith) are built for debugging your model, not for regulators. VouchRail sits next to them and writes the evidence chain.

Install

pnpm add @vouchrail/sdk      # TypeScript
pip  install vouchrail       # Python

Use

import Anthropic from '@anthropic-ai/sdk';
import { AuditLogger } from '@vouchrail/sdk';

const audit = new AuditLogger({
  systemId: 'hireflow-resume-screener',
  storage: { type: 'local', dir: './audit-logs' },
  signingKey: { kind: 'inline', secret: process.env.AUDIT_SIGNING_KEY! },
});

const anthropic = audit.wrap(new Anthropic(), {
  caseId: 'candidate-12345',
  promptTemplateId: 'resume-scoring-v3',
  promptTemplateVersion: '3.2.1',
  operatorId: 'system',
});

await anthropic.messages.create({
  /* … */
});

import os
from anthropic import Anthropic
from vouchrail import AuditLogger, InlineSigner, LocalStorageBackend, WrapContext

audit = AuditLogger(
    system_id="hireflow-resume-screener",
    storage=LocalStorageBackend(dir="./audit-logs"),
    signer=InlineSigner(os.environ["AUDIT_SIGNING_KEY"]),
)

anthropic = audit.wrap(Anthropic(), WrapContext(
    case_id="candidate-12345",
    prompt_template_id="resume-scoring-v3",
    prompt_template_version="3.2.1",
    operator_id="system",
))

anthropic.messages.create(...)

Async clients: audit.wrap_async(...). Decorator + context-manager surface: python/README.md.

Verify

Offline. No VouchRail cloud, no network.

vouchrail verify --storage-dir ./audit-logs --system-id hireflow-resume-screener
# ✔ Chain valid. N entries verified.

TS and Python write byte-identical chains — either CLI verifies either SDK's output. 16 conformance vectors in CI enforce it.

Docs

Full guides live in docs/:

• Quickstart · Architecture · Configuration reference
• Providers · Storage backends · PII redaction · Signing keys
• Schema + hash chain · CLI usage · Cross-language conformance

Why not just my observability tool

Four things observability stacks don't give you:

• Tamper-evidence. SHA-256 chain + external signing key. One altered record breaks the chain at that point — and you can prove which.
• Regulator-shaped schema. Model + prompt versions, input/output fingerprints, operator, human review, reason codes — fields a regulator can match to the regulation text.
• Your storage. Logs land in your S3 (or local disk). Vendor-hosted logs carry weaker evidentiary weight in court.
• GDPR-safe PII. Pseudonymize at write time; token store is deletable on request; the chain stays intact.

Status

Alpha. APIs may shift before v1.0. Working today:

• TS + Python SDKs at parity — wrap(), manual startCall/endCall, local + S3 storage (TS) / local (Python), inline + KMS signer, regex PII with in-memory + SQLite token stores.
• CLI in both languages: init / query / verify / export.
• Cross-language hash chain (RFC 8785 JCS + ECMA-262 NumberToString). 16 conformance vectors gate every build.

Not yet: LangChain / LangGraph adapters, hosted dashboard, ML-based PII, SOC 2.

Layout

• packages/schema — types, JCS, hash chain
• packages/sdk — AuditLogger, storage, PII, providers
• packages/cli — TS CLI
• python/ — Python SDK + CLI
• examples/ — HR screening, credit, healthcare triage
• conformance-vectors/ — cross-language byte-identity tests
• docs/ — topical guides (see above)
• legal/ — ToS, privacy, DPA, AUP templates (attorney review required before use)

Limits

Infrastructure, not legal advice. Not certified by any Notified Body. Evidentiary value depends on key custody, retention configuration, and periodic verification — see legal/limitations.md.

Dev

# Node
corepack enable && corepack prepare pnpm@8.15.9 --activate
pnpm install && pnpm build && pnpm test

# Python
cd python && pip install -e ".[dev]" && pytest

Contributing: CONTRIBUTING.md. Security: SECURITY.md. Conduct: CODE_OF_CONDUCT.md.

License

Apache-2.0. © 2026 VouchRail Contributors.