The log recalls each step in turn, the hour, the input, the score. It keeps the order of the march and never says what it was for.
The October hearing, having exhausted its four true denials in The Right Question to the Wrong Defendant, turns to the record. Here, at least, the institution has done its homework. The Succession handover file exists. The model version is documented, the prompt preserved, the retrieved evidence listed in order, and the determination timestamped to the second. The reviewer can walk the claimant through March step by step.
The central line of that walk-through reads something like this: the applicant's composite score fell below 0.63, and applications below 0.63 are declined.
It is a perfectly accurate sentence. The claimant listens to it, nods, and asks the same question again.
Sequence and grounds
A log answers the question what happened. It can do so with extraordinary precision, and institutions that have invested in evidence custody (as Attribution at the Moment of Action argued they must) can reconstruct a decision path in more detail than any human memory could hold.
The claimant is asking something else. Why was 0.63 the right place to draw the line, and why did that evidence count while this evidence did not? Why did falling below a number lead to this consequence, for this person, when a closer look was available? Those are questions about grounds, and a reconstruction contains none of them. The threshold appears in the log as a fact about the configuration. Its justification lives somewhere else, if it lives anywhere.
Authorization argued that accountability cannot be rebuilt from logs after the fact, because the answerable principal has to be present when the action commits. The difficulty here runs one level further down. Even a perfect log, kept by an independent custodian, with every principal named, still records sequence. The decision could be flawlessly reconstructed and remain entirely unexplained.
What the court asked for
European law arrived at this distinction by a slightly roundabout route. The GDPR gives a person subject to automated decision-making a right to "meaningful information about the logic involved." For years that phrase was argued over as though it asked for a technical disclosure: the model, the weights, the formula.
In Dun & Bradstreet Austria (Case C-203/22, February 2025), the Court of Justice of the European Union read it differently. A mobile operator had refused a customer a contract on the strength of an automated credit assessment, and the customer wanted to know why. The Court held that the controller had to explain the procedure and principles actually applied, in a concise, intelligible form, so that the person could understand which of their personal data had been used and how. Handing over a complex formula would not suffice. The explanation's purpose was to let the person understand the decision and contest it.
That is a reason, in the sense the claimant needs: something addressed to the affected person, built so they can take hold of it and push back. A reconstruction addressed to an engineer does not meet the standard, however complete it is. (The Court was careful to leave room for trade secrets, to be weighed by a court or authority case by case. Commercial confidentiality, it turns out, is also a kind of account that someone must give.)
The same distinction is now being argued from the top of the multilateral system. Tshilidzi Marwala, Rector of the United Nations University, took ten commitments grounded in the UN Charter and the Universal Declaration of Human Rights to recent discussions at the Pontifical Academy of Sciences, and his summary of them puts the problem in a single sentence: a decision "may be statistically effective yet lack legitimacy if no one can explain or contest it." His first commitment asks that rights be real for the person an algorithm affects, including a meaningful opportunity to challenge consequential decisions. A log can prove the effectiveness. Only a reason can carry the legitimacy.
The number and its reasons
Every threshold has a history. Someone chose 0.63, perhaps by calibrating against past outcomes, perhaps by fitting a target approval rate, perhaps by negotiating it with a regulator, perhaps by inheriting it from a predecessor system and never revisiting it. Each of those histories carries a different justification, and each would support a different conversation with the claimant.
A calibrated threshold can be defended with evidence about error rates. A threshold fitted to a target rate has to be defended as a policy choice about how many people the institution was prepared to decline, which is a far more awkward sentence to say to one of them. An inherited threshold has no defence at all except continuity, and continuity is precisely what Succession showed could not be assumed.
The log records the number in every case. It cannot tell the claimant which of those histories they are standing in.
Job's friends, arriving in the chapters ahead, will have a great deal to say about sequence. He sinned, they suggest; therefore he suffered. The order of events is tidy and the grounds are missing. Job's reply, in the King James rendering of 13:3, is the claimant's reply to every reconstruction offered in place of a reason: "Surely I would speak to the Almighty, and I desire to reason with God."
He wanted the reasoning. The record of what happened, he had already lived through.
Companions
- The opening question: The Right Question to the Wrong Defendant.
- Why logs alone cannot carry accountability: Attribution at the Moment of Action.
- The March file that makes reconstruction possible: The Harness Is What Stays.
- The ruling: CJEU, Case C-203/22, Dun & Bradstreet Austria, 27 February 2025.
- The human-rights framing: Tshilidzi Marwala, To Govern AI, Start With Human Rights.
- The fiction: the Book of Job, King James Version, Project Gutenberg ebook 10.
These notes come out of Sociable Systems, a practice that reads AI-shaped documents the way a hostile reviewer will, before a lender or a court finds the gap. The argument has an operational form: the Interim Protocol sets out four rules for AI use in environmental and social deliverables, covering disclosure at touch-point grain, evidence custody, the phrases no automated screening may settle, and a hostile read before anything ships. Free, and written to be cited or retired once institutional guidance arrives.