AI governance needs evidence

Article 12 of the EU AI Act requires high-risk AI systems to technically allow the automatic recording of events (logs) over their whole lifetime. The logs must support traceability: spotting situations where the system may present a risk or has been substantially modified, feeding the provider's post-market monitoring, and letting deployers monitor how the system runs.

Two further articles turn that capability into a duty to keep records. Providers (Article 19) and deployers (Article 26) must keep the logs under their control for at least six months, and longer where other EU or national law requires it.

What the Act does not do is say how to protect those logs from being edited. That gap matters more than it looks. A log that someone with admin access can quietly change is weak evidence in an audit or a dispute, however complete it is. TrustNotch closes the gap: every entry carries a signed receipt and a Bitcoin anchor that anyone can check without trusting us.

Using TrustNotch does not by itself make a system compliant. That depends on what you log, your role under the Act, and your legal context. What it gives your logs is the one property the law leaves to you: they can be shown not to have changed.

Most responsible AI and governance programs rest on a management system: policies, risk assessments, clear roles, and records that show the system runs as described. ISO/IEC 42001, the international standard for AI management systems, makes event logging an explicit control. Control A.6.2.8 asks you to decide at which stages of an AI system's life cycle event logs are recorded, and to keep logging on while the system is in use.

Frameworks like this say records must exist. They rarely say how anyone can trust them later. That is the layer TrustNotch fills: the audit trail under the program, not the program itself.

ISO/IEC 42001 certifies an organisation's management system, not individual tools. No product, TrustNotch included, can make you certified. What TrustNotch can do is produce records your auditors can check for themselves.

Most audit-log services store your records in a database they control. If a record needs to be shown unmodified, you must trust their assertion that nothing has changed — because no independent check exists.

In a regulatory audit or legal proceeding, this is a structural weakness. An auditor can ask for evidence that logs have not been tampered with after the fact — and "we say they haven't" is not evidence.

The same problem applies to AI systems themselves: an agent that can log its own actions to a store it influences has a credibility gap. Logs submitted to a service that returns signed, externally anchored receipts are harder to deny or revise.

TrustNotch does not decide what you record. That stays with you, your risk assessment and your role under the law. These are the events governance reviews tend to ask about first:

• Decisions: what the system concluded, with its inputs or a hash of them, the model or version, and when.
• Actions and tool calls: what an agent actually did, especially anything that changed data, spent money or contacted a person.
• Human oversight: approvals, overrides, and the moments a person stopped or corrected the system. The AI Act expects high-risk systems to support human oversight (Article 14), and these records show it happened.
• Changes to the system: model updates, configuration changes and new prompts, so later behaviour can be traced to the version that produced it.
• Incidents: errors, refusals, and anything flagged for review.

For personal data, send a salted hash rather than the content. The proof works the same, and the personal data never reaches us.

Tamper-evident logging and GDPR's right to erasure can seem at odds: if a log entry is cryptographically sealed, how can personal data be erased?

TrustNotch resolves this with per-entry erasure. You can erase the stored payload for any entry — removing it from TrustNotch's servers — while every proof for that entry and every sibling entry continues to verify. The Merkle tree is constructed over the hashes of content, not the content itself, so removing a payload doesn't invalidate any sibling's inclusion path.

This means you can satisfy a GDPR erasure request for a specific user's data without retroactively undermining the integrity of surrounding log entries.

TrustNotch's MCP server lets AI agents log their own actions directly — through the same Model Context Protocol interface they use for everything else. The agent calls submit_log, gets back a signed receipt, and can later call verify_log to confirm an entry is intact. The whole logging loop is in the agent's context, not a separate integration.

Because verification is independent of TrustNotch, an agent's logs can be audited by anyone — the operator, a regulator, or a third-party auditor — using the open-source verifier, without any involvement from TrustNotch.

• Not a governance platform. It does not keep your model inventory, run risk assessments or manage policies. It makes the records those processes produce verifiable.
• Not an observability tool. It is not built for dashboards, traces or debugging. It records the events you choose, so they can stand as evidence.
• Not a compliance guarantee. Using it does not by itself make a system compliant with the AI Act or certified under ISO/IEC 42001.