Working paper · Judgment & Answerability
The Licensing Problem in Predictive Maintenance
When Does a Prediction Become a Reason to Act?
Åke Elden · NLA University College, Oslo, Norway
Draft v5 — §2.1 and reference list verified against the publication record, July 2026; items marked [†] remain to be checked.
Abstract
Predictive maintenance research is organised around improving prediction. This paper argues that the binding constraint lies in the transition from prediction to intervention. I specify four conditions on a maintenance license — the entitlement to treat a predictive output as a ground for a given intervention — and argue their necessity by isolation. Three are grounding conditions: constitutive adequacy, entitlement transfer, and interventional discrimination. The fourth, answerability, is second-order, governing whether the first three can be held over time. The interventional condition binds asymmetrically: engineering knowledge supports replace versus do not replace, whereas predictive maintenance proposes replace now versus replace later. Four formal results follow, including that accuracy improvement resolves only the component of the intervention contrast the labelling scheme individuates, and that evidence accumulated by a control population buys a post-withdrawal extension depending on reserve intensity rather than duration. Reserve is a subscription, not an investment.
The claim
A prediction is a claim about what will happen. A maintenance decision is a commitment to make something happen. The first does not entail the second, and no accumulation of the first converts into the second. Something must license the passage — and that entitlement is what I call a maintenance license.
Contributions
Proposition 1
Accuracy improvement resolves only the component of the intervention contrast that the labelling scheme individuates. Where the binding constraint is constitutive, interventional, or institutional, better models have sharply limited licensing return.
Proposition 2–3
A license has a computable lifetime determined by process drift and the rate at which counterfactual evidence is regenerated — a lifetime independent of predictive accuracy. A programme can lose its entitlement to act while every monitored metric remains stable.
Proposition 4
Evidence accumulated by a reserved control population buys a post-withdrawal extension that scales with reserve intensity rather than duration. Reserve is a subscription, not an investment.
The full argument is withheld while the manuscript is under review. Available on request.