Pricing a signal before it arrives
Value of information is not a claim about how interesting a signal is. It is a price. Ronald A. Howard, formalising decision analysis in the mid-1960s, defined it by subtraction: the expected payoff of the best decision made with a signal, minus the expected payoff of the best decision made without it. Nothing about novelty enters the calculation. Nothing about volume enters it either. A signal that cannot change what you would do is worth exactly zero, however rich its content.
Perfect information sets the ceiling on this quantity — the expected value of perfect information, EVPI, the gap between deciding under uncertainty and deciding under omniscience. No signal can be worth more than that gap, because omniscience is the most any signal could deliver. Howard's own working examples came from petroleum exploration: a wildcat prospect with a modest chance of payoff, negative in expectation, where the question was never "is the seismic survey informative" but "does the survey change the drill/no-drill decision, and by how much is that change worth against the cost of running it." Wald's sequential analysis, built for wartime inspection, priced the decision of whether to sample once more. Blackwell's 1953 result gave the ordering beneath all of it: one information structure can be shown, for every possible decision-maker, to dominate another. Value of information is old, exact, and indifferent to how the signal was produced.
The lineage restated as a price schedule
The three generations of model discussed on this site are usually ranked by capability. Under value of information they are better read as three different purchasing arrangements for the same commodity: signal, priced against a decision.
A Large Language Model buys nothing at query time. Its information was priced once, at the training cutoff, and held fixed. Against any decision sensitive to what has changed since — a corroded flowline, a revised regulatory notice, a well that has drifted out of its expected pressure envelope — its expected value of information relative to the live world decays toward zero, because the model cannot know that the world has moved.
A Large World Model buys information at the moment of action, which is exactly where value of information is highest, since the decision is live and the signal is fresh. But the purchase is scoped to what the current scene emits and lapses when the scene ends. A camera on a wellhead sees that wellhead. It does not see the pipeline three kilometres downstream, and it forgets what it saw an hour ago unless something downstream was built to remember for it.
A Large Universe Model is the standing position: streams kept open past any single decision, carrying provenance so that the likelihood behind each belief — and therefore its price — remains computable when a decision finally goes live. This is the terminal rung not because intake stops mattering beyond it, but because "perfect information about every live stream" is the analytic ceiling of the quantity itself. There is no fourth class of evidence sitting past omniscience.
| generation | when it prices information | scope of what it can buy |
|---|---|---|
| Large Language Model | once, at cutoff | fixed corpus, decaying relevance |
| Large World Model | at the moment of action | the present scene, while observed |
| Large Universe Model | continuously, against live decisions | every open stream, with provenance |
The domain as test: oil and gas
Oil and gas is a reasonable place to test the claim because the decisions are expensive, the streams are numerous, and the cost of a wrong belief is measured in ruptured pipe, not lost engagement. Four streams matter to an integrity engineer on a producing asset: wellhead telemetry (pressure, temperature, flow, often at second- or minute-level resolution from SCADA), seismic surveys (episodic, expensive, spatially coarse), pipeline pressure and flow data along the gathering system, and regulatory notices — PHMSA advisories, state commission orders, the paperwork trail that changes what is legally permitted before it changes what is physically true.
Each of these has a different value-of-information profile because each feeds a different decision. A seismic survey re-run on a mature field is priced against the decision to drill an infill well or not; if the survey cannot move that call, Howard's zero applies regardless of how much the geology has evolved. Wellhead telemetry is priced against a faster, cheaper decision — choke adjustment, shut-in, workover scheduling — and because that decision recurs constantly, the value of a fresh reading is realised almost immediately or not at all.
The failure that makes the case
The characteristic failure in this domain is not exotic. An integrity signal — say, casing pressure on a well with a known annular anomaly, or corrosion-coupon data on a gathering line — is aggregated monthly, because that was the reporting cadence set when the system was designed, and because monthly summaries are cheap to store and easy to audit. The asset that carries that signal can fail in hours. A small leak at a compressor seal, or a stress crack propagating under cyclic pressure, does not wait for the end of the reporting period.
The value of information on that stream was never actually zero. It was mispriced. The decision it should have informed — shut in this well now, dispatch a crew to this segment today — is live at hourly resolution, and the signal was being delivered at monthly resolution, which means for twenty-nine days out of thirty the engineer was deciding under exactly the uncertainty the sensor was meant to remove. This is not a data problem. It is a decision-architecture problem wearing a data problem's clothes. The sensor was fine. The aggregation window destroyed the value of what it produced by delivering it after the decision it was meant to inform had already had to be made on other grounds.
An integrity engineer who has lived through a failure like this recognises the pattern immediately: the incident report shows the anomaly was "visible in the data" weeks before the rupture, and the honest response is that visibility was never the issue. The data existed. Its value of information against the shut-in decision was already zero by the time anyone with authority to act could see it, because the aggregation had already collapsed the six-hour signal that mattered into a thirty-day average that didn't.
Two objections worth taking seriously here
Value of information is only computable inside a fully specified decision model — known actions, known payoffs, known likelihoods. An operating field has none of these cleanly. Calling continuous telemetry "priced" borrows a rigour the setting cannot support, and unpriced ingestion is hoarding with a decision-theoretic gloss on it.
This is correct as stated, and it is worth conceding fully. Nobody has a complete payoff function for an integrity programme. What survives the concession is narrower and more useful: value of information is computed on slices, not on the whole field. This well's shut-in threshold. This segment's next scheduled pig run. This interim reading against a known corrosion-rate model. Those slices are exactly where continuous monitoring gets defended or killed in practice — an operator running a value-of-information calculation on whether hourly casing-pressure readings actually change the shut-in decision more often than daily readings would, and finding, in a specific case, that they do not, is doing the discipline correctly. Where no slice can be specified, the honest label for the ingestion is "unpriced," and it should be defended on option value or abandoned, not laundered through the vocabulary of decision analysis.
More information can lower payoffs. Alarm fatigue is a documented failure mode on SCADA systems: operators facing hundreds of low-priority alarms per shift stop responding to any of them, including the one that matters. The monotonicity value-of-information arguments assume — more signal, never worse — is a property of an ideal Bayesian agent with infinite attention, not a control room.
Also correct, and it identifies the actual constraint, which is not the volume of the stream but the non-negativity of value of information holding only for a single coherent decision-maker free to ignore a signal. A control room saturated with alarms is not free to ignore selectively; it is forced to ignore indiscriminately, which is a different thing and it does destroy value. The answer is not less intake. It is provenance and explicit ignore-rules — the mechanism that separates a properly built continuous-monitoring system from a firehose: knowing which sensor's reading is reliable enough, on which asset, to justify overriding a lower-priority alarm queue, and knowing it before the queue backs up rather than after the rupture.
Neither objection undoes the claim that continuous, provenance-carrying intake is the terminal position on this axis. Both sharpen what the terminal position has to include to be worth anything: a decision it is actually attached to, and a rule for when to stop looking at it.