Prompt and Configuration Drift Is Change Control With No Record

An AI feature's behaviour is determined by more than the model it calls. It depends on the instruction text, the parameters that control sampling, the tools it is allowed to invoke, the documents in whatever corpus it retrieves from, and the templates that assemble all of that into a request. Each of those is a live input, and in most organisations at least some of them are editable outside the software release process — in an admin console, a spreadsheet, a content management system, or a file a non-engineer was reasonably given permission to update. That arrangement is deliberate and mostly good; the people who know how the feature should behave are frequently not the people who deploy code. The problem is that the change-control machinery software engineering built over decades — version history, review, an audit trail, the ability to see exactly what changed and roll it back — usually does not extend to these surfaces. When behaviour changes for the worse, the first question is what changed, and the honest answer is often that nobody can reconstruct it. Retrieval corpora make this sharper because they drift without anyone taking an action at all. A grounded-answer system is only as correct as the documents behind it, and those documents are edited, superseded and archived by ordinary business processes that have no idea they are now part of a production system's behaviour. An answer that was right in March and wrong in August may involve no change to any prompt or model — just a policy page someone updated, correctly, for a different audience. Treating these inputs as configuration under version control is the fix, and it is less disruptive than it sounds: the editing interface can stay friendly while writing through to a versioned store, so that every change has an author, a timestamp, a diff and a revert. Add a small evaluation suite that runs on change, and the loop closes — behaviour shifts become attributable rather than mysterious. The organisations that skip this are not being careless. They are applying an accurate instinct that prompts are content, to a system in which content is now code.