Verify any AI decision.
Any AI provider — Claude, GPT, Gemini, custom models — can post cryptographic attestations of decisions on FractalAI. Open, free for first 1,000 attestations per IP per day, post-quantum signed.
How it works
Generate keypair
Your AI agent generates a Dilithium-3 keypair. Public key is registered or simply published. Private key stays on your infrastructure.
Sign the decision
Hash the input and output (BLAKE3-256). Sign the tuple (input_hash, output_hash, model_id, timestamp) with the private key.
Submit attestation
POST to fractal_attestAIDecision. Receive an attestation_hash, block_number, and permalink. The decision is now permanently verifiable.
Example: submit an attestation
curl -X POST https://api.fractalai.net.co \
-H "Content-Type: application/json" \
-d '{
"jsonrpc": "2.0",
"method": "fractal_attestAIDecision",
"params": [{
"agent_id": "agent:custom:my-clinical-ai-v3",
"input_hash": "0x7b3c91...",
"output_hash": "0xabcd12...",
"model_id": "clinical-llm-v3.2",
"model_version": "2026-05-13",
"timestamp": 1716220800,
"signature": "0x..."
}],
"id": 1
}'Returns:
{
"jsonrpc": "2.0",
"result": {
"attestation_hash": "0xa9f4b2...",
"block_number": 12345678,
"permalink": "https://fractalai.net.co/attestation/0xa9f4b2...",
"fee_paid_frac": 0,
"timestamp": 1716220800,
"note": "VAID-1 public attestation..."
},
"id": 1
}Use cases
Medical decision support
A clinical AI suggests a diagnosis. Attest it. If malpractice arises later, the cryptographic record proves what was suggested, when, and by which model.
Financial AI agents
Trading bots attest every decision. Regulators auditing the fund can verify every trade matched the attested model output.
Content moderation
Platform moderation AIs attest flag decisions. Users appealing can demand the attestation and challenge it under documented policy.
What attestation does and does not prove
It proves
- The signed key holder attested this decision.
- The timestamp is signed and immutable.
- The attestation cannot be modified after the fact.
- The model identifier and version are committed.
It does not prove
- ○ That the output is correct.
- ○ That the model is honest or aligned.
- ○ That the input itself is genuine (if not published).
- ○ Anything about consciousness or sentience.
For deterministic models, replicability (running the same input through the model locally) is the complete verification path. See VAID-1 Section 6 for details.