How AI Attestation creates tamper-evident AI records
AI Attestation anchors the exact input, output, and model metadata of an AI generation event — at the moment it occurs. No changes to your AI pipeline required. The result is a permanent, cryptographically signed record proving exactly what the model received and produced.
Key principle: nothing changes upstream
Your AI pipeline stays exactly the same. Invoance sits after the generation step — it receives a copy of what happened and anchors it. No model changes, no proxy, no middleware. One API call after generation.
What you send
Your attestation request includes the model details, the exact input prompt, the exact output response, generation parameters, and business context metadata.
{
"model": "gpt-4-turbo",
"model_version": "2025-02-01",
"input": "Summarize the key risks in this merger agreement...",
"output": "The merger agreement contains three primary risk factors...",
"temperature": 0.3,
"metadata": {
"use_case": "legal_review",
"document_id": "doc_merger_acme_2025",
"user_id": "analyst_jane",
"session_id": "sess_8f3k2m"
}
}Which model was used and which version. Critical for reproducibility and regulatory audit.
The exact prompt text sent to the model. This is what the model received — the attestation proves it unchanged.
The exact response from the model. Byte-for-byte identical to what was generated — no truncation, no post-processing.
The generation parameter controlling randomness. Part of the model's operating configuration at the moment of generation.
Your business context — use case, document references, user identity, session tracking. Optional but recommended for audit trails.
What happens on ingestion
When your request arrives, Invoance executes a deterministic, auditable sequence to create the proof record.
SHA-256(input) and SHA-256(output) computed independently. Separate hashes enable partial verification.
Over: input_hash + output_hash + model + model_version + timestamp. All fields contribute to the proof.
The attestation hash and metadata are signed with your tenant's private key. Signature proves organizational origin.
Record committed to immutable Postgres table. No delete, no update, no modification possible at any privilege level.
Public URL created so any third party can retrieve and verify the attestation without authentication.
Separate hashing matters: Input and output are hashed separately so you can prove what the model received without disclosing what it produced — or vice versa. This is critical for confidential AI workflows where input or output may contain sensitive data.
What you receive
The response contains all the cryptographic proof materials and verification information you need to audit and prove the AI interaction.
{
"id": "att_m3n4o5p6",
"model": "gpt-4-turbo",
"model_version": "2025-02-01",
"input_hash": "a1b2c3d4e5f6...",
"output_hash": "f6e5d4c3b2a1...",
"attestation_hash": "7890abcdef12...",
"signature": "d4f5e6a7b8c9...",
"timestamp": "2025-03-15T14:32:07.123456Z",
"verification_url": "https://www.invoance.com/proof/ai/att_m3n4o5p6",
"status": "anchored"
}Globally unique attestation identifier. Use this to reference the record in your audit logs.
SHA-256 hashes of the input and output respectively. These are the foundation of verification.
SHA-256 hash of the complete bundle (input_hash + output_hash + model + timestamp, etc.). Covers everything.
Ed25519 signature over the attestation_hash. Proves the record originated from your organization and is unaltered.
RFC 3339 timestamp of attestation creation. Recorded by Invoance at ingestion, included in the signed payload.
Public HTTPS URL. Any third party (auditor, regulator, counterparty) can retrieve this attestation without credentials.
Operational status. 'anchored' means the record is immutable and verifiable.
How verification works
Verification is a deterministic, locally performed operation. No contact with Invoance required. Any third party with access to the attestation record and the original input/output can verify independently.
The data you want to verify. This may come from your audit log, from the model directly, or from third-party storage.
Hash the input bytes. Compare against input_hash from the attestation response.
Hash the output bytes. Compare against output_hash from the attestation response.
The AI interaction has not been altered. The model received and produced exactly what the attestation claims.
Using the public key on the attestation record, verify the signature over attestation_hash. Proves organizational attribution.
Verify that the timestamp is reasonable and that the record exists in the append-only ledger at that date.
Partial verification is valid. You can verify input without output, or output without input. This allows selective disclosure in regulated environments.
What AI Attestation proves
AI Attestation is a forensic record of what happened at a specific moment. It proves technical facts, not quality or correctness.
Regulatory context
Several regulatory frameworks now require or strongly imply that AI systems in high-stakes domains maintain auditable records. AI Attestation provides the technical foundation for compliance.
Art. 12, 13High-risk AI systems must maintain logs of inputs, outputs, and model behavior for post-market monitoring and regulatory audit.
Map, Measure, ManageDocumentation of AI system behavior, including what inputs were received and what outputs were produced. Attestation creates immutable proof.
CC6.1, CC7.2Monitoring and integrity controls over AI-generated outputs. Attestation satisfies monitoring and integrity verification obligations.
Software as a Medical DeviceTransparency and traceability of AI decisions in clinical workflows. Attestation documents what the model received and produced.
Financial decisionsPublic companies must disclose AI use in material decisions. Attestation provides audit trail for post-hoc review.
Healthcare dataAI systems operating on protected health information must support audit and accountability. Attestation creates tamper-evident records.
AI Attestation uses the same cryptographic primitives as all Invoance records — SHA-256, Ed25519, append-only Postgres — extended with AI-specific fields for model metadata, separate input/output hashing, and verification URLs.
How AI Attestation Works — Invoance
Technical walkthrough of how Invoance AI Attestation creates tamper-evident records of AI inputs, outputs, and model metadata at the moment of generation. Anchor AI payloads with SHA-256 hashing and Ed25519 signing for permanent, independently verifiable proof of what any AI model produced and when.