ai-audit-trail 0.2.0

Tamper-evident AI pipeline audit trail — EU AI Act Art. 12 compliant Decision Receipts

ai-audit

The only standalone Python library that makes your AI pipeline EU AI Act compliant — cryptographically provable, no blockchain, no SaaS, no lock-in.

---

The problem

The EU AI Act (mandatory from August 2026) requires high-risk AI systems to keep tamper-evident logs proving every decision was made correctly. Most teams are solving this with home-grown logging — which is neither tamper-evident nor legally defensible.

Existing alternatives either require a blockchain (complex, expensive, slow) or a SaaS subscription (your audit data leaves your infrastructure).

ai-audit solves this with 3 lines of code.

---

Quickstart

pip install ai-audit-trail

from ai_audit import (
    AuditConfig, init_audit_config,
    ReceiptCollector, ReceiptStore, ReceiptAction,
    verify_chain, get_verify_key_hex,
    PiiConfig, PiiType,
)

# One-time setup
init_audit_config(AuditConfig.from_env())

# Wrap every AI request
store = ReceiptStore()
collector = ReceiptCollector(
    trace_id="req-abc123",
    tenant_id="acme",
    pii_config=PiiConfig(enabled_types=frozenset({PiiType.EMAIL, PiiType.IP})),
)
collector.set_input("What is our GDPR policy for [e-mail]?")
collector.add_check("safety", score=0.02, threshold=0.8, fired=False)
collector.set_output("Our GDPR policy is available at...")
collector.set_action(ReceiptAction.ALLOW)
collector.emit(store)
collector.cleanup()

# Verify tamper-evidence at any time
result = verify_chain(store.get_by_tenant("acme"), get_verify_key_hex())
assert result.valid  # Ed25519 + SHA-256 + hash-chain — all three verified

---

Why ai-audit wins

Feature	ai-audit	aegis-ledger-sdk	cortex-persist	ai-audit-sdk
Tamper-evident receipts	✅	✅	❌	✅
No blockchain required	✅	❌	✅	❌
Fully offline / air-gap	✅	❌	✅	❌
PII stripped before hash	✅	❌	❌	❌
EU AI Act Art. 9–18 report	✅	partial	❌	❌
SPRT confidence per article	✅	❌	❌	❌
Async-native	✅	❌	❌	❌
KMS / Vault / AWS KMS support	✅	❌	❌	partial
Redis Lua mode (1k+ req/s)	✅	❌	❌	❌
Standalone Python package	✅	❌	✅	❌
Zero external deps for PII	✅	❌	❌	❌

---

How it works

Every AI pipeline decision produces a Decision Receipt — a cryptographically sealed, hash-chained record that proves:

• What the model was asked (SHA-256 of normalised input)
• What it answered (SHA-256 of output)
• Which safety/quality checks ran and what they scored
• What action was taken (ALLOW / REJECT / ESCALATE / ...)
• Which model produced the answer
• That nobody tampered with the log (Ed25519 signature + hash-chain linkage)

Three-stage verification (< 0.1 ms per receipt)

Ed25519 signature  →  detects forgery (who signed it)
SHA-256 self-hash  →  detects corruption (content integrity)
Hash-chain link    →  detects insertions and deletions (ordering)

This is the same principle as a blockchain — without the blockchain overhead.

---

Installation

pip install ai-audit-trail

# Redis persistence (recommended for production)
pip install "ai-audit[redis]"

# Prometheus metrics
pip install "ai-audit[prometheus]"

# Everything
pip install "ai-audit[all]"

Requirements: Python 3.11+, no external services, works air-gapped.

---

PII-Redaction — GDPR Art. 17 compliant

Personal data is stripped before SHA-256 hashing. The stored hash is computed over the anonymised string — so even if your audit log is subpoenaed, it never contains raw PII.

from ai_audit import PiiConfig, PiiMode, PiiType

config = PiiConfig(
    enabled_types=frozenset({
        PiiType.EMAIL, PiiType.PHONE, PiiType.IP,
        PiiType.IBAN, PiiType.CREDIT_CARD,
    }),
    mode=PiiMode.REDACT,  # or HASH (deterministic SHA-256) or MASK (a***m)
)

collector = ReceiptCollector(tenant_id="acme", pii_config=config)
collector.set_input("Call [phone] or email [e-mail]")
# Stored hash = SHA-256("Call [PHONE] or email [EMAIL]")
# Raw PII never touches the audit log

Mode	Input	Stored as
REDACT	[e-mail]	[EMAIL]
HASH	[e-mail]	3d4e5f8a… (SHA-256)
MASK	[e-mail]	a***@c***.com

Supported types: EMAIL PHONE IP IBAN CREDIT_CARD CUSTOM (your regex)

Zero external dependencies — pure Python re + hashlib. Async via aobfuscate_text().

---

EU AI Act Compliance Reports

Generate legally-defensible audit reports in three formats — 100% offline, no internet required.

from ai_audit import build_compliance_summary, verify_chain, get_verify_key_hex
from ai_audit.report import ComplianceReportGenerator

receipts = store.get_by_tenant("acme")
chain = verify_chain(receipts, get_verify_key_hex())
summary = build_compliance_summary(receipts, chain_intact=chain.valid,
                                   verify_key_hex=get_verify_key_hex())

gen = ComplianceReportGenerator(summary, verify_key_hex=get_verify_key_hex())

gen.to_markdown()  # → Git, documentation portals
gen.to_json()      # → API consumers, automated pipelines
gen.to_html()      # → Self-contained HTML, works on air-gapped servers

Each report gives you:

EU AI Act Article	What it measures	How we measure it
Art. 9 — Risk Management	Safety gate firing rates	SPRT + guardrail activation
Art. 12 — Record-Keeping	Tamper-evident log integrity	Cryptographic hash-chain
Art. 13 — Transparency	Ongoing compliance confidence	SPRT confidence score
Art. 17 — Quality Management	System quality posture	SPRT certification status
Art. 18 — Automatic Logging	Real-time logging proof	ReceiptStore chain

Every report also includes the Ed25519 signing-key fingerprint — proving which system generated the receipts.

---

Continuous Compliance Certification (SPRT)

ai-audit runs a Sequential Probability Ratio Test on every request batch, giving you a live compliance status:

from ai_audit import build_compliance_summary

summary = build_compliance_summary(receipts, chain_intact=True)
print(summary.sprt_status)           # CERTIFIED | MONITORING | FLAGGED
print(summary.compliance_confidence) # 0.0 – 1.0
print(summary.check_fire_rates)      # {"safety": 0.02, "routing": 0.0}

Status	Meaning
CERTIFIED	Reject rate statistically below 5% — you're compliant
MONITORING	Insufficient data or inconclusive — watch this
FLAGGED	Reject rate exceeds 15% — investigation required

---

Production setup

Persistent signing key (mandatory in production)

python -c "import nacl.signing; print(nacl.signing.SigningKey.generate().encode().hex())"

init_audit_config(AuditConfig(
    is_production=True,
    signing_key_hex="your-64-char-hex-key",
))

If the private key is lost, historical chains can no longer be verified. Store it in your KMS.

KMS integration (Vault, AWS KMS, GCP KMS)

from ai_audit import KeyProvider, init_key_provider
import nacl.signing

class VaultKeyProvider(KeyProvider):
    def get_signing_key(self) -> nacl.signing.SigningKey:
        secret = vault_client.secrets.kv.read_secret("secret/ai-audit/key")
        return nacl.signing.SigningKey(bytes.fromhex(secret["data"]["key"]))

    def get_verify_key_hex(self) -> str:
        return self.get_signing_key().verify_key.encode().hex()

init_key_provider(VaultKeyProvider())

Redis persistence (high-throughput)

import redis
from ai_audit import ReceiptStore

r = redis.Redis(host="localhost", port=6379)

# Standard mode
store = ReceiptStore(redis_client=r, ttl=2_592_000)  # 30-day TTL

# Lua mode — single roundtrip, no connection-pool exhaustion at 1k+ req/s
store = ReceiptStore(redis_client=r, use_lua=True)

Prometheus monitoring

from prometheus_client import Counter
from ai_audit import verify_chain, get_verify_key_hex

CHAIN_BREAKS = Counter("audit_chain_breaks_total", "Hash-chain breaks", ["tenant_id"])

verify_chain(
    receipts,
    get_verify_key_hex(),
    on_chain_break=lambda tid: CHAIN_BREAKS.labels(tenant_id=tid).inc(),
)

---

Receipt schema

Every DecisionReceipt is a sealed, immutable record:

DecisionReceipt(
    receipt_id      = "a3f2...",      # UUID hex
    timestamp       = datetime(...),  # UTC
    trace_id        = "req-abc123",   # Correlation ID
    tenant_id       = "acme",         # Multi-tenant scope
    input_c14n      = "sha256:...",   # Hash of (anonymised) input
    output_hash     = "sha256:...",   # Hash of output
    checks          = [CheckRecord(name="safety", score=0.02, fired=False)],
    action          = ReceiptAction.ALLOW,
    model_id        = "gpt-4o",
    prev_receipt_hash = "sha256:...", # Hash-chain linkage
    receipt_hash    = "sha256:...",   # Self-hash
    signature       = "ed25519:...",  # Ed25519 signature
)

---

Environment variables

Variable	Default	Description
AI_AUDIT_ENV	development	Set to production to enforce signing key
AI_AUDIT_SIGNING_KEY	—	Hex-encoded Ed25519 private key
AI_AUDIT_RECEIPT_TTL	2592000	Redis TTL in seconds (30 days)

---

License

MIT — free for commercial use.

---

Built by S&S Connect · Feedback & issues welcome