H33-74 for AI Companies

What gets attested in ai companies

Each of the following operational events emits an H33-74 proof at the moment it occurs. The collection composes into the operational history that audit, compliance, and regulatory teams reconstruct from.

• Model inferences and the policy under which each was issued.
• Safety classifier decisions including thresholding rationale.
• RLHF reward model outputs and preference judgments.
• Content moderation determinations: allow, restrict, escalate, suspend.
• Agent actions: tool calls, decisions taken on behalf of users, escalations.
• Model lifecycle events: training run completion, evaluation outcome, deployment, deprecation.
• User-facing safety interactions including refusals and disclaimers.

Regulatory pressure that already justifies this

AI Companies is regulated under frameworks whose audit horizons now extend past the lifetime of the systems and vendors holding the evidence. H33-74 produces evidence that survives that horizon:

• EU AI Act. Article 12 logging plus Article 14 human oversight plus Article 17 quality management. See: EU AI Act crosswalk
• GDPR. Articles 5, 22, 30 — automated decision-making and recordkeeping. See: GDPR crosswalk
• NIS2. EU AI infrastructure providers. See: NIS2 crosswalk
• FedRAMP. AI services in federal cloud environments. See: FedRAMP crosswalk
• Voluntary commitments to the AI Safety Institute. US AI commitments under EO 14110 successors.

Scenarios this changes

Model provider changes the model underneath an application

Every prior decision the old model made remains independently verifiable. The application can answer regulator and customer questions about decisions made under the prior model without the model provider's cooperation.

Litigation requires evidence of safety classifier decisions

The proof of every classifier call, the threshold applied, the version of the classifier, and the disposition is produced. Verifiable without revealing model internals.

Audit of agent actions taken on behalf of an enterprise user

Each tool call, decision, escalation, and rationale is a separate proof. The enterprise verifies that the agent acted within authorized scope without trusting the AI vendor's log.

What survives versus what gets replaced

How to start

Three integration patterns, ordered by how invasive each is.

Pattern 1 — Receipt at the policy engine

The decision system (policy engine, approval workflow, AI scorer) emits an H33-74 proof for every decision at the moment it is made. No change to upstream systems. Auditors verify proofs directly.

Pattern 2 — Receipt at the system of record

Each system of record (treasury, claims, records, EHR) is wrapped so that committed transactions emit H33-74 proofs in addition to the existing audit log. Existing systems remain unchanged. Proofs become a parallel canonical record.

Pattern 3 — Receipt at the integration bus

Service mesh or integration bus intercepts cross-system events and emits H33-74 proofs. Highest reach, requires bus integration. Useful where systems-of-record cannot be modified.

Related