AI Attestation
Cryptographic verification for AI models, agents, and pipelines. Establish provenance, verify integrity, and prove compliance for AI artifacts.
Provenance vs Attestation vs Signing
Provenance
Tracks origin and history. Answers: Where did this come from? What transformations occurred?
DescriptiveSigning
Cryptographic approval. Answers: Who approved this? When was it signed?
Verification pointAttestation
Signing + provenance + policy + chain. Proves origin, integrity, compliance, and history with verifiable evidence.
Complete verificationWhat Can Be Attested
Model Weights
Full model files or delta updates with cryptographic binding to training lineage
Training Runs
Hyperparameters, compute resources, dataset references, and training metrics
Datasets
Training data with lineage tracking, filtering criteria, and preprocessing steps
Inference Logs
Model inputs, outputs, and decisions for audit and compliance verification
Agent Pipelines
Multi-model orchestration with policy bindings and execution traces
Software Artifacts
Build outputs, containers, and deployable packages with supply-chain attestation
When to Attest
Training Completion
Seal model weights and training metadata as baseline
Fine-tuning / Alignment
Attest delta changes with reference to base attestation
Pre-deployment
Verify chain integrity before production release
Runtime (periodic)
Generate inference attestations for audit trail
What Evidence Looks Like
Claims Status
Frequently Asked Questions
What is AI attestation?
AI attestation is the cryptographic verification of AI artifacts including models, training runs, datasets, and inference logs. It establishes who created an AI artifact, what inputs and processes were used, and provides tamper-evident proof that the artifact hasn't been modified since attestation.
What is the difference between AI provenance and AI attestation?
AI provenance tracks the history and origin of AI artifacts—where models came from, what data trained them, how they were transformed. AI attestation takes provenance further by cryptographically binding this history into verifiable, tamper-evident records. Provenance describes; attestation proves.
What is the difference between attestation and signing?
Signing proves who approved an artifact at a point in time. Attestation includes signing but adds structured metadata about what is being attested, policy compliance status, and linkage to the continuity chain. Attestation is signing with semantic context and verifiable history.
What AI artifacts can be attested?
Attested Intelligence supports attestation of: model weights (full or delta), training datasets and data lineage, training run metadata (hyperparameters, compute used), fine-tuning and RLHF records, inference logs and outputs, and agent execution traces.
When should AI artifacts be attested?
Key attestation points include: after training completion, after fine-tuning or alignment, before deployment to production, at regular intervals during inference (for audit), and after any model modification or update.
What evidence is included in an AI attestation?
AI attestations include: BLAKE2b-256 hash of the artifact, Ed25519 signature, RFC 3339 timestamp, artifact type classification, optional metadata (name, version, issuer), policy artifact reference if applicable, and chain linkage for continuity verification.
How does attestation help with AI regulation compliance?
Emerging AI regulations require transparency about model provenance and training data. Attestation provides the cryptographic evidence needed for compliance—verifiable records that can be audited by regulators, retained for liability purposes, and proven to third parties.
Can attestation prevent model poisoning?
Attestation doesn't prevent poisoning but enables detection. By establishing a verified baseline at training completion, any subsequent modification would break the attestation chain. Combined with supply-chain provenance, it creates defense-in-depth against model tampering.