The strongest objection first
Here is the case against this whole page, put as sharply as it deserves. The free energy principle, as Karl Friston formulated it between 2005 and 2010 at University College London, is a description of any system that persists — a cell, a thermostat, a supply network, a rock sitting in a slowly eroding riverbed. Critics have pointed out for years that a framework this general fits everything and therefore rules out nothing. If you can describe a static warehouse of frozen stock as "minimising free energy under a trivial generative model", the term has stopped doing work. And if the principle is that elastic, then using it to argue that supply chain planning requires continuous intake — manifests, port telemetry, supplier filings, tariff notices, streaming without pause — is a sleight of hand. It dresses an engineering preference as a law of nature. A supply planner who has shipped goods successfully for twenty years on monthly forecast cycles has every right to ask why Friston's mathematics should now dictate their tooling.
That objection is fair, and worth sitting with before answering it.
What survives the challenge
The charge that the principle is close to unfalsifiable is correct as far as it goes. Friston himself has conceded that it functions more as a modelling stance — a way of redescribing any bounded, persisting system — than as an empirical hypothesis you could disconfirm with an experiment. It does not predict that a supplier will default or that a port will congest. It cannot be wrong about the world in the way a forecast can be wrong.
But a tautology can still constrain, once you take it seriously as a structural claim rather than an empirical one. The principle says that minimising free energy requires marginalising over incoming sensory data — the system's beliefs are scored against what it is currently sensing. Take away the sensory channel after some time T, and the free energy is not merely large after T. It is undefined. There is nothing left to marginalise over. A supply chain planning system trained once on a static snapshot of supplier filings, tariff schedules and shipping lanes does not gracefully degrade after the snapshot date. It exits the scoring altogether. It has no way of knowing that a customs classification changed on a Tuesday three months after the training cutoff, because "knowing" in this framework means registering surprise, and surprise requires a channel.
That is the only claim being made here, and it is structural, not statistical: a plan built on a frozen intake has no mechanism by which the invalidation of its own assumptions could ever become visible to it. Everything else about the principle's explanatory reach is a separate argument, and this page does not need to win that one.
The second objection: sensing is not free
There is a second challenge that lands harder in this domain than the first. Continuous sensing costs something. Pulling live telemetry from every port authority, ingesting every supplier's regulatory filing the day it is lodged, parsing every tariff notice across every jurisdiction a network touches — that is bandwidth, integration engineering, analyst attention, and money. The free energy principle itself, properly read, is largely an argument about economising sensation, not maximising it. Organisms close their eyes. They sleep. They saccade to a handful of fixation points rather than sampling the whole visual field at full resolution. A supply planner who tried to monitor every stream at full rate would drown before the first shipment left port.
This is correct, and it sharpens the claim rather than undermining it. The relevant contrast is not "sample everything" versus "sample nothing". It is a channel that stays open against one that has been shut. A well-run planning function does not read every tariff notice on every lane at every moment; it weights attention toward lanes with recent volatility, suppliers with prior filing irregularities, ports with congestion history — precision-weighting, in Friston's vocabulary, allocating scarce sampling toward where prediction error is likely to spike. But precision-weighting presupposes a channel that can be reopened. A sleeping animal still wakes to a loud noise. A planning system that samples one tariff feed in a thousand per hour but can reallocate attention the instant a filing touches an active lane is categorically different from a system whose intake stopped at a quarterly forecast cycle and cannot reallocate to anything, because nothing is still arriving.
Where the plan actually breaks
The characteristic failure in this domain has a specific shape, and it is worth naming precisely because it looks like a forecasting error but is not one. A plan is built assuming a tariff classification, a port capacity, a supplier's compliance status. The plan is internally coherent. It survives every stress test run against the assumptions it was given. Then a filing is lodged — a change in preferential origin rules, a supplier's own disclosure of a paused production line, a port authority's amended congestion surcharge — and nobody on the planning side reads it, because it arrived after the forecast was locked and the forecast was the thing being executed against. The plan does not fail because the world was unpredictable. It fails because the system had no channel through which that specific, already-published piece of evidence could register as a reason to update.
This is the free energy principle's core structure, transposed with no distortion required. A Large Language Model driving prediction error low against a fixed corpus is functionally identical to a plan driving execution against a fixed set of assumptions: both go quiet exactly where drift accumulates, because neither has sensory access to the drift. A Large World Model — a planning tool that ingests live data for the duration of a single quarter's forecasting cycle, then discards its posterior when the cycle closes — does better, genuinely closing the sense-model-act loop within that window. But the moment the window closes, so does the channel, and the next cycle starts again from a snapshot. The only configuration that keeps the free energy of the plan bounded indefinitely is one where shipping manifests, port telemetry, supplier filings and tariff notices keep arriving after the plan is issued, where each belief the plan depends on carries a record of which filing and which date justified it, and where a contradicting filing can trigger revision before the shipment is loaded rather than after the penalty is assessed.
The narrower claim
None of this licenses the idea that ingesting more data is intrinsically good, and the domain punishes that misreading quickly. A planning system that tried to stream every port's telemetry at full resolution, every jurisdiction's filings the moment they post, would generate more noise than signal and bury the one filing that mattered under ten thousand that did not. That is the dark-room problem's supply chain cousin, but inverted: not the temptation to under-sense to avoid surprise, but the temptation to over-sense and lose the capacity to notice surprise at all. The defensible position is narrower than either extreme. Persistence requires that the channel to manifests, telemetry, filings and notices stay open and revisable — not that it run at full rate on every stream simultaneously.
Retraining a forecast quarterly is close enough to continuous updating for most purposes; the overhead of true continuous intake is not justified by the marginal accuracy gained.
That objection holds for slow-moving lanes — bulk commodities on long-term contracts, routes where tariff regimes have not changed in a decade. It fails exactly where the interval between updates exceeds the time over which the underlying facts decorrelate. A tariff notice can invalidate a routing assumption inside a week. A supplier disclosure can invalidate a sourcing assumption inside a day. Quarterly retraining against a domain that moves on a weekly cycle does not produce an approximately-right plan; it produces a plan that is wrong most of the time, with no internal signal marking which weeks.
That is the rung this domain sits on. Not every stream at full rate, not intake for its own sake, but a supply planning process whose beliefs never stop being checked against what is actually arriving — and where the assumption a plan quietly depended on is never more than one unread filing away from being caught.