prEN 18229-1 structures AI logging and claim checking

You must now document which AI model each workflow uses, log the evidence and reasoning behind every claim it makes, and set retention periods for each log category based on your own risk assessment under the EU AI Act and GDPR. The draft standard structures what you record; you choose how long to keep it and how to verify correctness.

The prompt is an analysis of 29 September 2026 of AI event logging and claim checking under the draft standard prEN 18229-1, which argues that the standard provides a framework for logging but leaves retention periods and hallucination detection to the organisation to define. The standard was published by the Slovenian Institute for Standardization in July 2026 and remains in draft phase. In our assessment, this means you cannot treat the standard as a complete solution: it tells you what fields to connect in a governance log, but the decision on how long to retain each category, and how to verify that AI outputs are correct, remains your own responsibility under existing data protection and AI regulation.

What fields must a governance log connect?

The draft standard treats logging as a distinct component within an AI trustworthiness framework. A usable event log proves more than that an answer appeared. A governance log for a sensitive workflow must connect the following:

• Model identity and version — which model generated the output and when it was deployed.
• Input and context — what the user asked and what information the system had access to.
• Generated claim — the specific assertion the model made, broken down to atomic, verifiable statements.
• Evidence and sources — the references, data or reasoning the model cited or should have cited.
• Verification outcome — whether each claim was checked, by what method, and what the result was.
• Human decision and escalation — whether a person reviewed the output, what they decided, and what action followed.

Decoupling these fields makes reconstruction possible. Without claim and decision context, a log remains a collection of individual answers without a demonstrable chain of accountability.

How long should you retain each log category?

The draft standard itself sets no universal retention period, and it would be incorrect to derive one from prEN 18229-1. The standard structures the event logging; the retention period per log category follows from other frameworks. The following factors determine how long you keep a specific log category:

1. Regulatory requirement — the EU AI Act and GDPR impose different retention duties depending on whether the output touched personal data or informed a high-risk decision.
2. Incident risk — outputs that could cause harm if wrong warrant longer retention to support investigation and audit.
3. Verification status — a claim you checked and verified may justify shorter retention than one you flagged as uncertain.
4. Domain sensitivity — healthcare, finance and legal advice typically require longer retention than general information.
5. Human oversight depth — if you kept a person in the loop, the decision context justifies retention long enough to demonstrate that oversight.

The practical choice is therefore not a single figure, but a classification per category. Performance measurements, model decisions, warnings and verification outcomes may be assigned different periods.

How should you separate generation from checking?

The academic literature shows that detection becomes more reliable when you split an answer into atomic, verifiable claims and test them separately. Research on claim decomposition and evidence assessment identifies a four-step process: break the answer into individual claims, find evidence for each one, assess whether the evidence supports it, and locate precisely where hallucination occurs. This makes it possible to know not only that an answer is wrong, but also where.

A critical operational requirement emerges from research on multi-agent checking: the system that verifies claims must test them independently without taking the original answer as its starting point. Systems where a model assesses its own output show confirmation bias. The draft standard structures event logging, but organisations must use methods that document how generation and checking are separated.

A second warning comes from research on hallucination detection in realistic dialogue. Standard language models as detectors still perform insufficiently when applied to real user interactions rather than artificial benchmarks. Domain risk and human escalation therefore remain part of the process; you cannot rely on automated checking alone.

What governance steps must follow verification?

A verified output is not an end point but a versionable governance object. The following steps connect to the log fields from prEN 18229-1 and to the claim checking process:

1. Decompose the output into verifiable claims — break the model's answer into atomic statements that can be tested separately.
2. Check each claim against independent evidence — verify references and reasoning without bias toward the original answer.
3. Document the verification outcome — record which claims passed, which failed, and where hallucination occurred.
4. Escalate uncertain or failed claims to a person — do not publish or act on unverified outputs without human review.
5. Version the output and its log — treat the verified answer as a governance record that can be audited and updated.
6. Retain the full chain — keep model identity, input, claim, evidence, verification and decision together for the period your risk assessment requires.

Checking the underlying references is part of this process. The common thread is that the combination of the European logging standard and recent detection research shifts the starting point from "the AI has answered" to a demonstrable chain of event, evidence, judgement and governance action.

Automated tooling can structure your logs, flag claims for checking, and track which outputs have been verified. What remains your own professional judgement is the risk classification that determines retention, the choice of which claims warrant human review, and the decision to act on an output only after verification is complete.