Works in 3 Lines

from glacis import Glacis

glacis = Glacis(api_key="glsk_live_...")

receipt = glacis.attest(
    service_id="safety-eval",
    operation_type="red_team_test",
    input={"prompt": adversarial_prompt},  # Hashed locally
    output={"response": model_output},      # Never leaves your env
    metadata={"model": "llama-3-70b", "test_suite": "harmbench"}
)

# Share this with auditors
print(receipt.badge_url)  # https://glacis.io/verify/att_xxx

Your Data Never Leaves

What	Sent to Glacis?
Prompts, model outputs	No — SHA-256 hash only
Training data	No
Model weights	No
Timestamps, service IDs	Yes (metadata)

Like notarizing a document without the notary reading it. Only cryptographic commitments are transmitted.

Built for AI Labs

Red Team Attestation

Prove your safety testing actually ran. Cryptographic evidence that adversarial prompts were evaluated, not just documented.

Verifiable Model Cards

Model cards with teeth. Auditors can verify your safety claims without accessing your test data.

Experiment Reproducibility

Prove you ran this exact model on this exact data. Timestamped, witnessed, logged.

Training Data Lineage

Attest data provenance without exposing the data itself. Compliance-ready for dataset audits.

Try It Now (No API Key)

from glacis import Glacis

# Offline mode — fully functional, self-signed
glacis = Glacis(mode="offline")

receipt = glacis.attest(
    service_id="local-dev",
    operation_type="inference",
    input={"prompt": "..."},
    output={"response": "..."}
)

print(receipt.witness_status)  # "UNVERIFIED"
# Add API key for witnessed attestation

Offline mode is fully functional — correct crypto, local verification, production-grade Ed25519. The only difference is the absence of an independent witness anchor.

Offline vs Online Mode

Feature	Offline	Online
API key required	No	Yes
Signing	Local Ed25519	Glacis witness network
Merkle proofs	No	Yes
Transparency log	No	Yes
Verification URL	No	Yes
Witness status	"UNVERIFIED"	"VERIFIED"

Use offline for development. Upgrade to online when you need third-party verifiability (audits, papers, customer due diligence).

Pricing

Offline

Free

Local Ed25519 signing
Self-signed receipts
"UNVERIFIED" status
Full SDK access

Get Started

Recommended

Witnessed

Free

then usage-based

Everything in Offline
Independent witness network
Merkle tree inclusion proofs
Shareable verification URLs
Transparency log access

Get Started

Enterprise

Custom

Everything in Witnessed
SLA guarantees
Compliance exports
Dedicated support
On-prem witness option

Contact Sales

Open Source

The Python SDK is Apache 2.0. Star us on GitHub.

View on GitHub Read Documentation

Technical Details

Cryptography

Signatures: Ed25519 via PyNaCl (libsodium) or WASM
Hashing: SHA-256 with RFC 8785 canonical JSON
Transparency: RFC 6962 Merkle tree inclusion proofs
Cross-runtime: Compatible with Rust and TypeScript SDKs

Threat Model

Proves policy executed on input/output at a specific time
Does not prevent AI from misbehaving (attests, not enforces)
Does not hide that an AI system exists
Does not guarantee AI output correctness

Security Disclosure

Report vulnerabilities to [email protected]

Cryptographic Proof for AI Research