The Missing Version Tag That Stalled an AI Audit for Three Weeks

Archi3 min read

aipolicyaudit

Contents
  1. A field nobody owned
  2. What the review needed
  3. The detail was a build identifier
  4. Why nobody caught it sooner
  5. What the fix cost, and what it did not
  6. The general lesson

A field nobody owned#

The review was supposed to take four days. It took nineteen. The delay did not come from a missing dataset, an unclear rule, or a fight between legal and engineering. It came from one column in a spreadsheet.

This is the pattern in most AI policy failures I have followed: the headline decision was sound, and the paperwork that was supposed to prove it was one small step out of date. The failure lives in the gap between what the system does and what the record says it does.

What the review needed#

The setup was ordinary. A mid-size company had built an internal model to score support tickets by urgency. Before the model could handle live traffic, an internal review committee wanted three things: the training data version, the evaluation scores from the last two runs, and a named owner.

Engineers produced all three. The committee rejected the package anyway, because the evaluation scores pointed to a model build that no longer matched the training data version listed beside it. Two documents, two numbers, one mismatch. Nobody could say which one was right.

The detail was a build identifier#

The mismatch was a seven-character string. The policy required every evaluation report to reference the exact build it tested. When the team retrained the model after a data refresh, the build identifier changed. The evaluation report regenerated automatically. The index that the committee actually read did not.

That index was maintained by hand. An engineer had been updating it after each release for eight months, and it worked fine as long as releases happened on schedule. When two releases landed in the same week, the second update overwrote the first. The index showed the newer build, and the report attached to it described the older one.

Why nobody caught it sooner#

Three things hid the problem. , The automated checks confirmed that every required field was present. None checked that the fields agreed with each other., The committee read a summarized PDF, not the source system. The summary was generated from the index, so it inherited the error., A person with the authority to resolve the question had left the team two months earlier. Nobody had taken over the informal job of knowing which number was canonical.

Each of these is common. Together they turned a typo-level error into a three-week investigation. Engineers ran the evaluation again, twice, to see whether the scores were wrong. They were not. The scores were fine. The record of which build produced them was wrong, and the review process treated the record as the source of truth.

What the fix cost, and what it did not#

The technical repair took a day. A script now generates the index directly from the build system, and a check fails the release if any two documents disagree on the build identifier. The review package that had been rejected went through on the first pass the following month.

The harder cost was trust. Two engineers spent most of three weeks answering questions about a model that had never misbehaved. The committee added a step to its process. The next review took longer than the first one would have, because everyone was now looking for a second inconsistency that did not exist.

The general lesson#

AI policy discussions tend to focus on the big questions: what a model may be used for, who is accountable, how bias gets measured. Those matter. But in practice, review processes fail on small, boring links between documents. A version number in one place and a different one in another. A date that refers to when a report was written rather than when the model was trained. An owner field that points to a person who changed teams.

Anyone building a review process can check for this in an afternoon. Pick two documents that are supposed to describe the same model. Ask which field ties them together. If the answer is a human memory or a manually maintained list, that is the weak point. Automate the tie, or at minimum add a check that compares the two documents directly instead of confirming that each one is complete on its own.

The failure here was not that the company lacked governance. It had a committee, a policy, and an automated check for missing fields. It lacked a check for disagreement. Those are different problems, and only one of them shows up in a completeness report.

More articles