Every AI decision should be independently auditable years later, even if the model, the provider, and the wrapping application have all changed.
H33-74 produces one cryptographically verifiable proof per ai decisions action. The collection of proofs is the operational history. The history survives the systems that produced it.
What gets attested in ai decisions
Each of the following AI events emits an H33-74 proof at the moment of execution:
- Model inference: the prompt commitment, the model version, the temperature and sampling parameters, the output commitment, the policy applied.
- Safety classifier output: the input commitment, the classifier version, the threshold applied, the classification result.
- RLHF judgment: the comparison pair, the rater identity, the preference, the rationale.
- Content moderation call: the content commitment, the policy in effect, the disposition (allow, restrict, escalate).
- Agent action: the tool invoked, the parameters committed, the user authorization, the result.
- Model lifecycle: training run completion, evaluation outcomes, deployment, deprecation, all as separate proofs.
Workflow
1. Decision produced
The AI system makes a decision (inference, classification, moderation call, agent action). The decision context — input, model version, policy, output — is hashed to a 32-byte commitment.
2. Receipt emitted
The H33-74 substrate produces a 74-byte receipt signed under three post-quantum families. The receipt records the computation type, the commitment, and the timestamp.
3. Anchor scheduled
The receipt enters the anchor queue. Batched anchoring posts a Merkle root over many receipts to the chain or chains the operator anchors on.
4. Verification on demand
An auditor verifying any AI decision later requests the receipt and the Merkle proof. The verifier confirms the three PQ signatures and the anchor commitment. No AI vendor cooperation required.
Regulatory anchoring
This use case lives under specific regulatory frameworks. H33-74 produces evidence that maps to each:
- EU AI Act Article 12. Automatic logging of AI system activity. Each H33-74 receipt is one logging entry. See the crosswalk.
- EU AI Act Article 14. Human oversight events emit receipts. Override decisions produce verifiable evidence.
- GDPR Article 22. Automated decisions producing legal or significant effects emit receipts including the decision logic available to the data subject. See the crosswalk.
- SOX 404. If AI is part of ICFR, each AI decision is part of the audit trail. See the crosswalk.
What survives infrastructure change
Survives
Every ai decisions action your operation produced — independently verifiable forever.
Gets replaced as lifecycle work
The systems, the vendor SaaS contracts, the chains you anchored to.
The ai decisions evidence is chain-portable. When the underlying systems change, the ai decisions record carries forward as PQ-signed proofs anchored to whichever chain is appropriate at the time.
Related