The Licensing Problem in Predictive Maintenance

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.

The four conditions

C1 — Constitutive adequacy

The failure modes individuated by the labelling scheme must include the modes the intervention addresses, and the relation between the labelling threshold and the functional failure of concern must be documented and defensible.

C2 — Entitlement transfer

The deployment situation must lie inside an envelope specified in advance — including operating context and maintenance regime, not only input distribution — with the mechanisms assumed shared between validation and deployment stated.

C3 — Interventional discrimination

The sign of the intervention value must be determined by evidence that does not reduce to observational conditioning under the policy already in force. Established engineering knowledge supports replace versus do not replace; it rarely supports replace now versus replace later.

C4 — Answerability

There must exist an agent with authority to enact and to withhold the action, access at decision time to the grounds for C1–C3 as documented, and exposure to the consequences. C4 is second-order: it governs whether an entitlement grounded by C1–C3 can be held over time.

Formal results

• 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.

• Corollary 2

  Lowering the alarm threshold raises the evidence-generation cost of the license in proportion to the additional degradation range admitted. The threshold is not a free hyperparameter; it sets the evidential burden the programme must carry.

Structure of the argument

1. § 1

   Introduction

   Licenses are situation-indexed, they expire independently of model accuracy, and they require a bearer. The argument is neither against predictive maintenance nor in favour of human judgment over automation: time-based and reactive regimes operate under licenses too, and usually weaker ones.

2. § 2

   The closure thesis

   Prediction plus preferences is taken to close the question of action. Closure is not false; it is correct only under four rarely stated conditions — which are the licensing conditions themselves, appearing in the guise of assumptions about a formula.

3. § 3

   Four gaps

   The constitutive gap (models predict a proxy, not failure), the entitlement gap (warrant is a relation, not a property of a model), the interventional gap (observational conditioning is not an interventional contrast), and the answerability gap (a defeasible license requires a defeater-detection mechanism).

4. § 4

   The maintenance license

   The predicate L(f, a, σ, r) holds when an organisation is entitled to treat a predictive output as a ground for performing an action on a situation under the responsibility of an agent. Metrics reach only the first argument place. Necessity is argued by isolation.

5. § 5–6

   Four formal results

   Accuracy improvement resolves only the component of the intervention contrast the labelling scheme individuates; licenses have a computable lifetime governed by process drift; and reserved non-intervention buys a post-withdrawal extension scaling with reserve intensity rather than duration.

6. § 7

   What licensing is not

   Not model validation (validation compares outputs to the predictand; licensing relates the predictand to the action). Not explainability (neither necessary nor sufficient). Not threshold tuning (the threshold sets an ongoing evidential burden, not a hyperparameter).

7. § 8–10

   Objections, operational schema, conclusion

   Answers to the MLOps, decision-theory, prescriptive-maintenance, and tacit-expertise objections; then a register in which the four conditions, their grounds, and their expiry are recorded and owned.

Positioning

The prescriptive-maintenance literature has named the gap: predictive capability translates poorly into prescriptive capability, because associative models cannot support interventional claims. What it has not established is what the conditions on legitimate action are, how they relate to predictive accuracy, and what sustaining them costs.

Where this sits in the programme