When an AI output is disputed, the team needs to reconstruct the conditions in which it was produced. Versioning turns an anonymous piece of text into a controlled artefact with a known provenance.
Assign a unique run ID
Link every output to its timestamp, user, workflow and environment. The ID should follow the artefact even when it is copied into another system.
Record exact versions
Retain the model identifier, prompt template, parameters, tool definitions and post-processing code. A general model name is insufficient when the provider changes its version.
Capture inputs securely
Keep hashes or controlled snapshots of source material, depending on its sensitivity. Do not copy personal data into logs unnecessarily.
Link human changes
Record whether the output was approved, corrected or rejected, and which version was published. The initial generation is not always the final artefact.
Test reproducibility
Check whether you can explain an earlier result and repeat the workflow. Randomness permits variations, not the loss of the process itself.
This framework is an original editorial methodology developed by DigitalNow.