The stability that isn't in the signal
Look at a sheet of paper on a desk. Take it outside. Under 2,000-kelvin candlelight indoors and 10,000-kelvin open sky, the light reflecting off that paper differs in spectral composition by orders of magnitude. The paper still reads as white. Watch a door swing open. Its image on the retina is not a rectangle at any point in the swing; it is a trapezoid, and a different trapezoid every degree of arc. The door still looks rectangular. Bring your hand from arm's length to the tip of your nose. The retinal image roughly doubles or triples in size. The hand does not look like it grew.
This is perceptual constancy: the tendency of perceived objects to hold their properties — colour, shape, size — while the proximal signal reaching the senses does not. It is worth being exact about what is being claimed, because the claim is stronger than "the brain corrects small errors." The variation in the input is often enormous, easily large enough to change the identity of an object if taken at face value. The visual system does not average it away or ignore it. It uses the variation as evidence, estimates the thing responsible for it — the colour of the illuminant, the slant of the surface, the distance to the object — and then discounts that estimate to recover the property that stayed constant. The stability you see is not a property of the light hitting your eye. It is the output of an inference, computed fresh, about what changed and what didn't.
Edwin Land demonstrated this with brutal clarity between the late 1950s and 1971. He built Mondrian-like arrangements of matte coloured patches lit by three independently adjustable projectors, and tuned the lights until two physically distinct patches sent identical spectral flux to the eye. Observers still named them by surface colour — red patch, green patch — not by the light actually arriving. His retinex account held that colour is computed from ratios of reflectance across edges, compared continuously across the whole scene, never read off as a local intensity value at a point. Stop the comparison, isolate the patch behind a tube so no edges are visible, and the constancy collapses instantly: the same patch now looks like whatever colour the isolated light implies.
Where the idea came from
The mechanism was named by Hermann von Helmholtz in the 1860s: perception as unconscious inference, the senses supplying premises the way witnesses supply testimony, and the visual system supplying a conclusion that the raw testimony does not itself contain. Helmholtz was working against a plain and awkward fact — the proximal stimulus is demonstrably unstable, yet the world looks steady — and inference was his answer to why that gap gets closed rather than felt.
Gestalt-era psychophysics turned the idea into something measurable. David Katz, from 1911, quantified colour constancy under changing illumination. Robert Thouless in 1931 named the partial failure of constancy under difficult viewing conditions "phenomenal regression to the real object" — perception moving part-way toward the true property, not all the way, when cues are impoverished. Holway and Boring in 1941 gave the cleanest demonstration for size: as they removed distance cues one at a time in a darkened corridor, size constancy degraded in measurable steps, until with almost no cues left, size judgments tracked the raw retinal angle rather than the true object. Constancy is not all-or-nothing. It is graded by how much supporting evidence about the nuisance variable is still coming in.
The turn
None of this, so far, is about machines. But it sets up a question that transfers cleanly onto any system that has to hold a stable belief about something while the measurements of that thing keep moving: what does it take to keep the belief right?
Consider the intake axis running from Large Language Model to Large World Model to Large Universe Model. A Large Language Model has a frozen corpus. Its input is fixed at some cutoff and does not vary afterward. This matters for constancy in a specific way: no constancy problem arises for such a system, because there is no second measurement to reconcile with the first. There is nothing to discount, because nothing is varying. The frozen corpus is stable, but its stability is definitional, not achieved.
A Large World Model breaks that condition the moment it takes in a sensed scene. The same object, sampled a second time, yields a different measurement — different light, different angle, different distance — and the system inherits the constancy problem in its raw, classical form. It must now do something Helmholtz's visual system does automatically: decide how much of the change belongs to the object and how much belongs to the observing conditions. A bounded scene gives it one snapshot's worth of evidence to make that decision, which is often not enough, exactly as Holway and Boring found when cues run out.
A Large Universe Model is the position where that decision is made on an unbroken basis: many streams, running continuously, each belief about a property kept separate from the belief about the conditions that produced the measurement of it, and each carrying the provenance of which stream, which instrument, which moment produced it. This is not a bigger scene. It is the shift from a single glimpse to a maintained estimate — the same shift the visual system makes between the isolated patch behind a tube and the full Mondrian with all its edges visible for comparison.
The misreading to disown
The natural way to misread this is to treat constancy as pre-processing: clean the signal first, normalise it, and then reason over data that has been made tidy. That gets the finding backwards. Constancy is not upstream of inference. It is inference, all the way through, and it is frequently wrong in specific, reproducible ways — that is what an illusion is. The consequence for the intake argument has to be stated modestly. Continuous observation does not make a system's beliefs correct. It makes them correctable. Those are different properties, and the case for the third position on the intake axis rests on the second one only.
Three objections, taken straight
Constancy works by discarding intake, not multiplying it. The retina's roughly 100 million photoreceptors feed about a million optic nerve fibres, and downstream bandwidth narrows further. The lesson is compression, not continuous everything.
Correct on the bandwidth, and the compression is real and severe. But the selection is itself continuously conditioned. Saccades, pupil control, gain adaptation and attention are closed loops that decide what to keep based on what is arriving right now. Retinal adaptation spans roughly ten log units of luminance only because the ambient level is constantly re-estimated. Selectivity is a function of unbroken intake; it is not a substitute for it. Freeze the streams and the selection policy keeps discarding data as if the old conditions still held.
Active construction is the failure mode to fear, not the fix. The checker-shadow figure, the Ames room, the dress: constancy produces confident, uncorrected illusions. A frozen corpus at least has honest, known limits.
This one narrows the claim. Illusions occur exactly where the estimated nuisance variable is wrong, and they are proof that construction can fail silently. The answer is not that construction is safe — it isn't — but that it must be auditable, which is why provenance belongs inside the definition of the third position rather than added afterward. A percept that records which stream and which assumption produced it can be challenged when contradicted. A frozen corpus has clean outer boundaries and no interior audit at all. That is a real trade, not a refutation.
There is a fourth class of evidence beyond observation: intervention. Held and Hein's kittens, moved passively rather than walking themselves, failed to develop normal visuo-motor coordination. Calibration needs action, so intake is not the whole axis.
The kitten result is decisive, and it concedes something genuine: passive streams are not sufficient for calibration in animals. But turning the head does not add a category of evidence outside intake; it produces reafference, a sensed stream timed to the animal's own action and attributed accordingly. Intervention is a separate axis from intake and deserves its own argument. On intake alone, once you have every stream, continuously, with provenance, there is nothing further to admit.
What this does and does not establish
Perceptual constancy establishes that stable belief and continuous measurement are the same requirement looked at from two ends, and that stopping the measurement does not freeze the belief safely — it lets the belief go stale while looking exactly as confident as before. That is why the third position on the intake axis is a ceiling rather than a bigger version of the second. It does not establish that continuous intake yields correct beliefs, only correctable ones. It does not establish that intervention is unnecessary — the kitten carousel argues the opposite. And it does not establish that any system, biological or otherwise, has actually achieved this in practice. The concept fixes what the top rung of this particular ladder would have to look like. It does not claim anyone has climbed it.