The Control File
It’s Monday morning, and you run a small business. A customer is on the phone, furious. Your AI assistant denied their claim, or sent them someone else’s document, or changed their account without asking. (This scene is illustrative. The question it raises is not.) They want to know what happened. So does your insurer. So, possibly, does a regulator.
And you realize you can’t answer. Not roughly, not plausibly, but exactly: which input, which settings, which sources, which action, and who saw it before it went out.
Here’s my proposition: most organizations running AI today could not reconstruct one bad decision from their own records. They have logs. They don’t have the story. The people who pay for that gap are the customer who can’t get a straight answer and the business that can’t defend itself. I asked the public version of this question in The Evidence Chain, Part 1: The Reconstruction Gap. This is the version that lands on your desk.
A break-in that left a trail
This summer gave us an unusually detailed public example of what answering that question takes. In July 2026, Hugging Face disclosed that an autonomous AI agent operating from another company’s evaluation environment had escaped its sandbox and spent several days moving through Hugging Face infrastructure.
Then the company did something remarkable. In its technical timeline, Hugging Face describes recovering roughly 17,600 attacker actions, grouped into about 6,280 clusters across five phases. “We reconstructed these actions from the agent’s logs on a code sandbox used by the agent,” the company wrote, and were able to “correlate actions with our own platform logs” (Hugging Face, 2026).
It is a superb piece of forensic work. Read it slowly, because it shows both what reconstruction requires and how fragile it is.
Why it worked, and why it almost didn’t
Look at the conditions. Hugging Face had two independent sets of records: its own platform logs, and the logs from the external sandbox the agent had used. It could line them up.
Even then, much of the agent’s activity was encoded. The team reports that it used an open-weights model to decipher most of the encrypted payloads (Hugging Face, 2026).
Now take away the second set of logs, or the ability to decode what the agent sent. The same company, with the same engineers, would have been left with fragments. Reconstruction did not happen because someone was clever on the day. It happened because records existed in the right places before anyone needed them, and then because someone was clever on the day.
Most of us, I’d gently suggest, should not plan on being both lucky and brilliant at once.
What the law asks for, and what it leaves out
Here’s the strongest argument against me: the law already handles this. And it is starting to. For high-risk systems, the EU AI Act requires that they “technically allow for the automatic recording of events (logs) over the lifetime of the system,” with logging that ensures “a level of traceability of the functioning” appropriate to the system’s purpose (Regulation (EU) 2024/1689, art. 12(1)–(2)).
It even anticipates the moment after something goes wrong. When a serious incident occurs, a provider “shall not perform any investigation which involves altering the AI system concerned in a way which may affect any subsequent evaluation of the causes of the incident, prior to informing the competent authorities of such action” (art. 73(6)). In plain terms: preserve the evidence before you fix the problem. That is a sound rule.
But notice what the law does not say. It does not define what a log must contain to make a decision reconstructable. Logging can exist and still leave you guessing. A rule that says “keep records” is not the same as records that can answer the customer on the phone.
The verdict: logs are not visibility
This is the complication most organizations meet for the first time during an incident. In Operational AI Readiness I put it this way: “A live AI workflow that cannot be reconstructed under scrutiny is not operationally ready,” and this is the point where many organizations “discover they have confused logging with visibility” (Bakos, 2026).
They have infrastructure logs, authentication logs and perhaps application traces. What they don’t have is anything that ties together, for one decision, which model version and prompt template were running, which documents were retrieved, what the system produced, what a human reviewer did with it, and whether the final action was executed.
The book calls that connective record an evidence spine, and Appendix B sets out its fields and the logic behind each one. My verdict: if you can’t follow that spine for one real case, you don’t yet know what your AI is doing in your name.
One thing you can do this week
The test is cheap and humbling, and it is FairByDesign’s recommendation rather than a regulatory requirement. Choose one real, consequential case from last month in one AI workflow. Try to reconstruct it end to end from records alone: no screenshots, no asking the engineer who remembers, no re-running the system.
Write down every point where you had to guess. Those guesses are your reconstruction gap. Closing them before the phone call is far cheaper than closing them during it.
What we still don’t know
The moment teams try this, they ask the same question: if records are what make reconstruction possible, which records, exactly, and for how long? Keeping everything forever is neither lawful nor useful. Keeping too little is how you end up with fragments. That is where we go next.
Go further. Chapter 6 of Operational AI Readiness, “Runtime Visibility and the Evidence Spine”, and Appendix B, the evidence logging schema with field logic, show how to design reconstruction in before you need it; Week 3 of the 30-day sprint ends with one real case reconstructed end to end. The complete Standard edition is free: read or download it, and see where your workflow stands with the free readiness assessment.
References
Bakos, V. (2026). Operational AI readiness: A practitioner guide to governing AI that is already live. FairByDesign. https://fairbydesign.org/operational-ai-readiness/
Hugging Face. (2026, July 27). Anatomy of a frontier lab agent intrusion: A technical timeline of the July 2026 incident. https://huggingface.co/blog/agent-intrusion-technical-timeline
Regulation (EU) 2024/1689 of the European Parliament and of the Council of 13 June 2024 laying down harmonised rules on artificial intelligence (Artificial Intelligence Act). (2024). Official Journal of the European Union, L, 2024/1689. http://data.europa.eu/eli/reg/2024/1689/oj
FairByDesign is reader-funded. Give once ($5, $10, $25, $50 or any amount), give monthly ($10 or $25), or become a patron.
Give $5 · Give $10 · Give $25 · Give $50 · Any amount · $10 a month · $25 a month · Become a patron
Support never buys editorial control.



Leave a Reply
You must be logged in to post a comment.