The meeting that already happened
An economist walks into a rate-setting meeting carrying the latest print of a consumer price index, a labour-flows release, a credit-aggregate table and a survey of market expectations. All four will be revised. The inflation figure released this month for last month is not the number that will sit in the historical series six months from now — it is a provisional estimate, built on a partial sample, that will be revised as more returns come in, sometimes by enough to change the sign of the month-on-month move. The decision gets made today, on today's numbers, for a world that will not fully report itself for another two quarters.
This is not a scandal. It is the ordinary condition of monetary policy. But it is worth asking, carefully, what kind of knowledge a rate decision represents when it is built on data destined to change under it.
Russell's clock at the rate-setting table
Bertrand Russell's example is a stopped clock. You glance at it once, at the one moment its frozen hands happen to match the true time. You form a true belief. You are not wrong. But you do not know, because the method that produced your belief — reading a broken clock — would have given you the identical answer at any other moment, true or false. Nothing about your method tracks the fact. Philosophers call the gap veritic luck: true, even justified, but true by coincidence rather than by a process sensitive to how things actually are.
Central bank forecasting has its own stopped clocks, and they are not metaphorical. A committee that raises rates in response to a strong labour print is, in a strict sense, responding to a number that will be revised. If the revision confirms the story, the decision looks prescient. If the revision reverses it — a payrolls figure cut by a wide margin two months later, a credit aggregate restated after a reporting error at a handful of large lenders — the decision looks wrong in hindsight, even though nothing about the committee's reasoning was defective at the time. Both outcomes can happen to the same method. That is the signature of luck, not of skill: the belief would have been formed identically whether or not the eventual fact matched it.
The three generations of intake-bearing systems sit at different points on this problem, and central banking makes the differences concrete rather than abstract.
A frozen corpus — the Large Language Model's condition — is the economist's stopped clock with no update mechanism at all: a briefing document written before a release, unable to move when the release lands. A bounded, sensed scene — the Large World Model's condition — is closer to a live dashboard during the meeting: prices, spreads and survey data streaming in, sensitive to the room's inputs, but blind the moment the meeting ends and blind to anything the dashboard was never wired to show, such as informal credit conditions among unlisted lenders. The claim under test here is that neither position, however improved, escapes the revision problem on its own — and that only a third position, one that keeps every stream open indefinitely and treats each figure as provisional with a recorded lineage, closes the gap that revision keeps reopening. Call that third position the Large Universe Model: not a forecasting product but the argued endpoint of an axis, the one where intake never stops and every belief carries the trace of the release that produced it and the revision that may overturn it.
Two positions, honestly stated
Set the case for that endpoint against its strongest rival.
The reliabilist defence of ordinary practice. A committee does not need counterfactual sensitivity to every future revision. It needs a process that is right often enough, calibrated against known error bands, to justify acting under uncertainty. Inflation nowcasts built from provisional data have a measurable track record; their error distributions are published; policy is set with fan charts precisely because everyone concedes the point estimate will move. On this view, demanding that a rate decision "track" the eventual, fully-revised truth is a philosopher's standard imported into a domain that already has a better one: known reliability, quantified, published, revisited. Insisting on modal sensitivity smuggles in a commitment the profession never signed up to and does not need.
The whole apparatus of fan charts and error bands exists precisely because we know the number will move. Calling that veritic luck is describing weather forecasting as astrology because tomorrow sometimes surprises the model.
There is real force here. For most of what a price index reports — that a barrel of oil got more expensive, that shelter costs rose faster than wages — the direction is stable across revisions even when the magnitude is not. A committee is not usually fooled by a stopped clock; it is navigating a clock that runs slightly fast or slow, and it knows the drift rate.
The anti-luck challenge to that comfort. The trouble is at the edges, not the centre. A stable error band describes a stationary regime: the same kind of revision, of the same rough size, happening for the same reasons, release after release. It says nothing about the release where the regime itself changes — where a labour market classification is redefined, where a credit aggregate suddenly excludes a category of shadow lending that had been quietly inflating it, where a currency peg's credibility collapses between one data cut and the next. In those moments the historical error band is itself stale, computed from a world that no longer obtains, and the committee has no internal signal that this release is different from the ninety before it. That is exactly the condition under which a frozen or narrowly-scened method is right by accident when it is right at all: the fan chart looked calibrated because nothing regime-breaking had happened yet, not because the method had any grip on regime breaks as such.
What continuous intake buys, and what it does not
Two objections deserve a direct answer here, because central banking is where they bite hardest.
The first: continuous intake does not remove luck, it manufactures a new supply of it. A high-frequency credit aggregate updated daily can be miscalibrated at the source — a data vendor double-counting interbank exposures, a survey panel that has quietly become unrepresentative as smaller lenders drop out of the sample. A stream that is wrong every day is not improved by being fresh; if anything it is more dangerous, because freshness invites the false confidence that recency equals warrant. A committee that trusts the live number more than the revised one, purely because it arrived yesterday, has not escaped the stopped clock. It has adopted a faster one, still not synchronised to the fact.
This is why the terminal position on the intake axis cannot be defined as intake alone. It has to be intake plus provenance plus revisability. Provenance means each figure in front of the committee carries a visible trail — which survey wave, which sample size, which known exclusions, which prior revisions to the same series and by how much. Revisability means the system is built to retract, not just append: last month's inflation print sits alongside a flag showing it superseded an earlier estimate and may itself be superseded again, rather than being presented as settled fact until the next release quietly swaps it out. Redundancy across independent streams — cross-checking a labour flow figure against payroll tax receipts, checking a credit aggregate against bank balance-sheet filings collected on a different cycle and a different definition — turns a lucky coincidence between two numbers into an actual corroboration. None of this abolishes the possibility of being wrong. It makes the wrongness visible while it is still correctable, instead of visible only in the revision six months later.
The second objection: knowledge is the wrong standard for a policy meeting anyway. What a committee needs is a defensible decision under a stated error tolerance, not a philosopher's certification that its beliefs meet a sensitivity condition. This is largely right, and it should be conceded without hedging. Most meetings are not regime-break meetings. Most revisions are within the historical band, and a committee that waited for unrevisable certainty would never move rates at all — inaction is also a decision, made under exactly the same uncertainty. The objection fails only at the specific point where the historical band stops describing the present: and the only available warning of that point is a stream that is still running, cross-checked, and dated, rather than a single provisional print treated as if it were the fact itself.
The narrowed claim
The honest resolution is not that a rate-setting committee is permanently the victim of Russell's stopped clock, nor that continuous, provenance-tagged intake would make policy immune to revision surprise. Revision is structural to how price indices and labour statistics are collected; no amount of streaming removes the lag between an economic event and its full statistical settlement. What continuous, provenanced intake changes is narrower and more defensible: it converts an undetectable regime break into a detectable divergence between independent streams, and it converts a silently superseded figure into a visibly flagged one. That is not the abolition of epistemic luck in central banking. It is the reduction of one specific kind of it — the kind where a committee cannot tell, from the document in front of it, that the document has already stopped being true.