01 The problem
Every measurement is a purchase
An inspection takes the asset out of service for as long as it lasts. A high-fidelity simulation burns days of compute before it answers. A destructive test consumes the very part it was meant to qualify. In each case the act of finding something out has a price, and that price is paid before anyone knows whether the answer will change what happens next.
Engineering practice has always known this, which is why inspection intervals are negotiated rather than chosen freely. What is missing is a way to state the problem so that a machine can reason about it. Simulation environments almost universally hand observation over for free: the agent sees the state, or a noisy version of it, at no cost and at every step. A policy trained that way is never asked the question a maintenance engineer answers every week, which is whether this particular measurement is worth what it costs.
02 Why it matters
Maintenance is one of the largest recurring costs, and much of it is spent finding out
Maintenance is where the price of information is paid most visibly, because so much of it is organised around finding out rather than fixing: the calendar inspection that finds nothing, the borescope that confirms the engine is fine, the ground test that clears an aircraft to fly. The figures below are for aviation, where the accounting is public and unusually careful, but the structure repeats wherever an asset is expensive to open and expensive to leave closed.
Maintenance
Inspection calendars are set conservatively because the alternative is a failure in service. Every interval that could safely have been longer is money spent on an answer that was already known.
Design
A design loop queries oracles of different fidelity and different price, from a coarse surrogate to a full simulation. Which one to call, and when, is the same decision wearing different clothes.
Qualification
A destructive test is the extreme case: the measurement and the part are the same object, so the information is bought with the thing it describes.
03 Why it is hard
What a measurement is worth depends on what you would otherwise have done
Information has no value in itself. Its value is the change it makes to a decision, which means it cannot be priced without a model of the decision it feeds, and that model has to hold while the system it describes is degrading.
A measurement that changes nothing is waste
If the same action follows whatever the answer turns out to be, the measurement was worth nothing, however accurate it was. Deciding this in advance requires reasoning about an answer not yet received.
Free data is rarely neutral
Sensors that report continuously and for free often carry an installation bias that no amount of reading them reveals. Detecting it requires the paid measurement whose necessity is in question.
One budget, two demands
Money spent finding out is money not spent acting, and the two compete inside the same budget. A policy that is careful about repairs and careless about tests has not solved the problem.
Benchmarks that give sight away
An environment that hands over the state for free cannot measure how well a policy spends its budget for reducing ignorance, because in that environment the budget does not exist.
04 Where it stands
A contract that prices observation, and a result that cut both ways
Our work on this is COIN, a proposed contract for environments in which every action declares whether it changes the system or changes what is known about it, and carries a price either way. Two engineering environments implement it, one where nothing useful is free and one where free sensors are rich but biased, and a frozen protocol benchmarks both.
This is a research preview, v0.4, and the benchmark refused as much as it granted: paying for measurements wins outright where information is scarce, and wins no cost advantage at all where free data is already good. The note states the protocol, the numbers and what is deliberately not claimed.