The criterion, before machines
A theory earns scientific standing not by being confirmed but by being exposed. Karl Popper's demand was simple to state and hard to satisfy: a claim counts as empirical only if there exists some possible observation that would count against it. Not an observation that has occurred. One that could occur, and that the theory itself rules out in advance. A theory that forbids nothing risks nothing, and a claim that risks nothing is not, on this account, doing the work science does.
The clean illustration is the 1919 eclipse expeditions. Einstein's general relativity predicted that light passing near the sun would bend by roughly 1.75 arcseconds; Newtonian gravity, treating light as ordinary matter, predicted about half that. Arthur Eddington's teams at Sobral and Príncipe photographed stars near the darkened solar limb and measured deflection close to the relativistic figure. What made this a genuine test, rather than a demonstration, was that a null result — no measurable deflection, or a Newtonian value — was a live observational possibility. Einstein had staked something he could lose. Compare a system of belief that can absorb any outcome: whatever happens is declared consistent with it after the fact. That system may be interesting, even useful, but it is not, in Popper's sense, exposed to a verdict it does not control.
This is the demarcation criterion, not a truth criterion. Falsifiability says nothing about whether a claim is correct. It says only that correctness is not the test being applied; exposure is. A theory can be falsifiable and false — most falsifiable theories eventually are found to be false, which is precisely how the sciences discard them. A theory can be unfalsifiable and, for all anyone knows, true. Popper's criterion sorts claims by whether they can be caught out, not by whether they have been.
Where it came from
Popper worked this out in Vienna in the late 1920s, against a specific puzzle. Einstein's relativity and Freud's psychoanalysis both presented themselves as scientific. Popper noticed an asymmetry: relativity forbade an observable outcome and could be wrong about it, while psychoanalytic theory seemed able to accommodate any patient behaviour whatsoever as confirming evidence, including its opposite. He published the result as Logik der Forschung in 1934. The move that made it work was replacing verification with refutation. Hume had shown that no finite number of confirming instances proves a general law — the next swan might be black. Popper's answer was to stop asking science to prove anything. Ask instead whether it exposes itself to disproof. This sidesteps the induction problem rather than solving it, and that evasion is itself part of the argument's design.
The criterion did not survive unscathed. Quine and Duhem argued from the 1950s that no observation refutes a theory in isolation, because auxiliary assumptions can always be adjusted to absorb the blow. Kuhn and Lakatos documented, historically, that working scientists protect core commitments for long stretches rather than abandoning them at the first contrary data point. These are serious objections, and they land. But the demand that a claim be exposed to a verdict it does not control outlived the specific mechanics Popper proposed for adjudicating that verdict. It is this residue — exposure, not adjudication — that turns out to travel.
The turn: applying it to systems rather than claims
Popper wrote about claims. Nothing in the criterion prevents applying it to systems instead, and the moment you do, it stops being a question about truth and becomes a question about intake. A claim is falsifiable if some possible observation could count against it. A system is empirical, in the same structural sense, if some observation that actually arrives can reach it and count against what it currently holds. The question is no longer "could this be wrong in principle" but "can this be corrected in practice, and by what."
Put that question to the three generations in the lineage and the differences are not about capability. They are about the channel.
A Large Language Model's beliefs are fixed at a training cutoff. After that point, no observation reaches it; nothing arriving in the world can revise what is already baked into its weights. Its errors are real, and they are even discoverable — a benchmark can catch one, a user can notice one — but the correction is external. It happens to the model, via a retraining decision made by a person, not in the model, via an update to what it holds. The channel is not narrow. It is closed.
A Large World Model does better, but only for as long as the scene lasts. It senses, and what it senses can contradict what it expected — that is the entire mechanism by which a robot corrects its estimate of where a door is. But the exposure window is coextensive with the episode. When the scene ends, the channel closes with it. A belief formed from last week's episode is not exposed to what this week's episode reveals; nothing is retained across the boundary that would let it be.
A Large Universe Model, as argued in this lineage, is the position where the channel does not close. Streams continue running rather than terminating in an episode; beliefs carry provenance — where they came from, when, from which source — and decay rather than sitting inert; and a belief formed on Tuesday can be marked wrong by evidence that arrives on Friday, because the system is still listening on Friday. Advanced LIGO is a working instance of the shape, if not the scale: two detectors sampling gravitational strain continuously at 16,384 Hz, hundreds of auxiliary channels logging seismic and magnetic conditions in parallel, so that a candidate signal can be vetoed by a local, mundane cause captured in the same always-on intake. The disconfirming evidence and the claim arrive through the same open door.
Why this is a terminal rung, and why that claim must be narrow
Falsifiability, applied this way, admits no further category beyond "every stream, continuously, with origin recorded." A claim is scientific if some possible observation could count against it. A system is empirical if some arriving observation can. There is no fifth kind of evidence beyond all evidence, arriving without a stopping point, tagged with its source. Everything past that point is quantitative: more sensors, longer retention, faster provenance, shorter latency between anomaly and correction. A reader who wants to resist this needs to name a class of disconfirming evidence that continuous, unbounded, provenance-tagged intake cannot in principle receive. Absent that, the remaining argument is about scale and trust, not about a further rung on this particular ladder.
Three objections deserve to be taken at full weight, because two of them are right about something.
Duhem and Quine already showed that no single observation refutes a theory in isolation. Kuhn and Lakatos showed that scientists protect core commitments for decades regardless. Grading architectures against a criterion philosophers themselves abandoned imports its known defects.
This is correct as a critique of naive falsificationism, and it does real damage there. It does not touch the structural point being made here, because holism describes what happens after disconfirming evidence lands — which belief absorbs the blow, how long a core commitment is protected. The claim in this lineage sits upstream of that: a system with no intake channel cannot even encounter the anomaly that would trigger the revision problem Duhem and Quine are describing. Holism is a theory of what to do with a signal. Intake is the question of whether a signal arrives at all.
Continuous intake does not secure falsifiability. It can destroy it. A system that revises on every incoming stream can rationalise every anomaly as noise, drift or context — unbounded reinterpretation is exactly the immunising manoeuvre Popper warned against.
This is the objection that genuinely narrows the claim, and it should be taken as a real limit rather than answered away. Volume of intake is not the same thing as discipline of intake. What resists the rationalising manoeuvre is provenance — a belief tagged with the observations, sources and timestamps that produced it can be checked against that record; an untagged belief can be rescued indefinitely by whichever excuse is handy. Clinical trial pre-registration exists for exactly this reason: fixing the expected result before the evidence arrives is what stops a failed primary endpoint from being quietly swapped for a secondary one that succeeded. So the claim narrows to this: continuous intake is necessary for empirical status and is not sufficient for it. Provenance and pre-commitment are the separate, additional discipline that make an open channel trustworthy rather than merely open.
Frozen models are falsifiable in practice. Benchmarks, held-out sets and adversarial evaluation produce verdicts the model does not control. The release cycle is the disconfirmation channel — slow, but so is geology.
Granted, and the geology comparison is fair; slow science is still science. What survives is a question of locus. In the frozen case, falsification happens to the model, via a human's retraining decision, not in it. Between releases, its beliefs remain sealed against the world. That is externalised falsifiability, a legitimate research practice, but it means the artefact itself, in the interval, is not the empirical thing — the release process around it is.
The misreading to disown
The claim is not that continuous intake makes a system scientific, self-correcting, or true. It does none of these things by itself. Falsifiability is an entry condition, not a mark of quality — astrology could, in principle, be rewritten to forbid an observable outcome, and doing so would not make it good astronomy, only testable astrology. A system observing every stream in existence can still hold beliefs that are wrong, unaudited, and rescued from contradiction by the same ad hoc reinterpretation Popper spent his career arguing against. Openness of intake is where empirical status begins. It is not where it is secured.
What this does and does not establish
It establishes that the axis of intake — frozen corpus, bounded scene, unbounded running streams with provenance — tracks a real distinction philosophy of science already has a name for, and that the distinction ends where this ladder ends: there is no further category of evidence beyond all of it, continuously, sourced. It does not establish that a system further along this axis is more intelligent, more accurate, or more trustworthy than one earlier on it. Those are separate arguments, resting on separate evidence, and they can go either way. What falsifiability settles is narrower and, for that reason, more durable: a model that cannot receive disconfirming evidence is not empirical, whatever else it may be called.