The Systems Thinker on the fog is the schema without a mirror
1. Fog as equalization, not absence
Stated: “Fog is not the absence of light. Fog is the equalization of light.”
Formalized: Take the variable to be the direction a photon arrives from. In clear air, radiance is sharply peaked by direction, so a photon from the northeast tells you something about the northeast. Multiple scattering in dense fog pushes that distribution toward uniform. The light doesn’t go away. What goes away is the mutual information between arrival direction and source location, I(direction; source). Navigation is inference on that quantity, so navigation fails.
Evaluation: Holds, with one physical correction. A single droplet scatters mostly forward. Isotropy belongs to the multiply scattered field, not to any single scattering event. And real fog seldom reaches the limit: some vertical gradient usually remains, brighter above than below. This matters later.
2. Maximum entropy as maximum silence
Formalized: sisuon uses “information” in the negentropy sense, log N − H(p). That is the divergence of p from the uniform distribution, and it is zero at maximum entropy.
Evaluation: Partially holds, and the reader should know which sense is in use. In Shannon’s own accounting, a uniform source carries maximum information per symbol, because every outcome is maximally surprising. What it lacks is redundancy: nothing about it can be predicted, compressed, or learned. So the fog is full of surprise and empty of anything learnable. sisuon’s own phrase, “the silence of everything happening at once,” is the precise version. The claim holds for learnable structure and is inverted for surprisal. The argument only needs the first sense.
3. The schema as a partition that hides by working
Formalized: A schema is a coarse-graining: a partition of the state space into cells (“this, not that”) plus a prior over those cells. The line “the sorted world is evidence for the schema, not evidence of it” has an exact reading. Evidence expressed in a partition’s categories can compare hypotheses stated inside that partition. It cannot compare the partition with alternatives, because the alternatives’ categories never show up in the data.
Evaluation: Holds, and it matches predictive processing. A generative model that predicts well produces little prediction error. Little error draws little attention, so the model becomes transparent: it is experienced as the world rather than as a model. “The organ of sight rather than the filter on the lens” describes that transparency accurately.
4. The fog test
Formalized: Under a uniform distribution, every partition into k equal cells has the same entropy, log k, and the same predictive value, zero. No way of carving is privileged. If a schema still sorts confidently under those conditions, its output is coming from its prior.
Evaluation: This is the document’s strongest claim, and it has a close empirical counterpart. Reverse-correlation methods feed a perceptual system pure noise and then average the samples it labeled “signal.” The average shows the system’s internal template. In Gosselin and Schyns’s work on “superstitious perception,” observers told to find a letter in pure noise produced classification images that looked like that letter, though the letter was never in the stimulus. The Ganzfeld effect is literal fog: a uniform visual field fills with percepts after a few minutes. Predictive processing supplies the mechanism: when the input stops being reliable, priors take over. sisuon’s claim that fog can reveal a schema is more than a metaphor. It describes a working method.
5. “The schema was always projecting”
Evaluation: Here the structure leaks. The fog test shows that a schema can produce output without structure in the input. It does not show that the schema produced its output the same way when the air was clear. In a world with low entropy, partitions are not equal: some compress better, predict better, and carry more information about future states. A schema that learns from prediction error would not keep sorting the same way in fog. Its confidence would flatten toward uniform. So “always projecting” holds for a frozen schema and fails for a learning one.
There is a version that holds, and it is stronger than it first looks. Even real asymmetries in the world underdetermine the schema. Goodman’s “grue” problem shows the same evidence confirming incompatible predicates equally well. No-free-lunch results show that no learner beats another without assumptions matched to the problem. Minimum description length depends on which description language you choose, though the invariance theorem limits that dependence to an additive constant. That gives sisuon’s absurdity a precise shape. The ground is chosen, the data don’t force the choice, and the world limits how much the choice can matter. “Not nihilism — the schema isn’t wrong. It works.” The formal picture agrees, and adds that the absurdity is real but bounded.
6. The zeroth step: fog → accident
Formalized: This is spontaneous symmetry breaking. A ferromagnet cooled below its Curie point has to pick a magnetization direction. Every direction is equally available, a fluctuation picks one, and the choice locks in. The Pólya urn is the cleanest model. Start with one red ball and one blue ball. Draw one, then put it back with a copy. The red proportion converges almost surely, and the value it converges to is itself uniformly distributed on [0,1]. The outcome is stable, real, and not necessary in any sense.
Evaluation: Holds, including sisuon’s correction of their own claim: “The accident didn’t break the fog. The fog produced the accident.” In symmetry breaking, the symmetry survives across the ensemble of possible runs, and only each individual run breaks it. That is the relation the corrected sentence describes.
The chain as quoted is missing two things, though.
- An amplifier. At true equilibrium, fluctuations die out instead of recurring. For “a microstate that… happened to recur” to harden, something has to reinforce it: positive feedback, a system sitting at an instability, or increasing returns like the urn’s copy rule. Fog at equilibrium produces accidents that vanish. Only fog near an instability produces accidents that become syntax. In the chain quoted here, the “recurrence” arrow is quietly doing that work, and it should be named.
- Degeneracy. The magnet’s possible states are equally good. Schemas in a structured world are not. The genetic code is partly a frozen accident and partly shaped by chemistry. “Every schema is absurd” is true of the schema’s frozen-accident part, not of the whole schema. The fog reveals that part.
Concept map
Nodes:
- F: the input field, with entropy H(F)
- S: the schema (partition plus prior)
- W: the sorted world (S applied to F)
- R: the amplifier (recurrence or reinforcement)
- Att: the observer’s attention
Edges:
- F → S through R: how a schema starts (fluctuation plus amplification)
- S × F → W: sorting
- W → S: confirmation. This is a reinforcing loop, and nothing balances it against an outside standard. That loop is the circularity sisuon describes.
- low prediction error → Att withdraws from S: the schema becomes transparent
- H(F) → maximum: W decouples from F and S is exposed (the fog test)
Boundary: sisuon draws the boundary around S and W and puts F outside, as the fog that “doesn’t care.” The key structural question is whether F → S is an ongoing edge (learning) or happened only at the start (genesis). sisuon’s argument is fully valid in the second case and only partly valid in the first.
The three hidings
The taxonomy of loom, meridian, and schema maps cleanly onto three different operations:
| Hiding | Formal role | Detection |
|---|---|---|
| Loom | a constraint on which states can be reached (a boundary condition) | inspect the boundary |
| Meridian | structure in the variables the partition discards | increase resolution |
| Schema | the partition itself | test it with noise as input |
These are genuinely different operations, so the taxonomy holds. My reading of the first two terms depends on this document’s summaries of the loom and meridian notes, not on those notes in full.
The complication of the modularity note (“the contract declares its surface but not its groundlessness — and no contract can”) also holds in a precise sense. An interface specifies behavior within the partition it chose. It has no way to express the partitions it didn’t choose. The link to the symmetry note gets sharper under the symmetry-breaking reading: a broken state presents itself as if it were unbroken.
The prescription as temperature
“Hold the schema while knowing it’s absurd” has a natural formal version: a finite temperature, where temperature sets how much probability goes to alternatives other than the current best. At zero temperature the schema is frozen. At infinite temperature it dissolves into fog. sisuon is asking for a temperature floor, enough that alternatives keep some probability. Then revision can happen “not because something went wrong,” through something like entropy regularization or the exploration term in active inference, rather than through an error signal.
Summary assessment
The strongest structural claim is that fog detects schemas. It holds, and there is a well-established method that already puts it into practice.
The weakest is the jump from “groundless in fog” to “always groundless.” To make it precise, split the schema into two parts: the part the data constrain and the degenerate part, the frozen accident. sisuon’s absurdity holds for the second part and is bounded by the first. The split can be measured. Run the schema on fog, and the structure left in its output is its prior. Run it on clear input and subtract. The difference is what the world contributed, and it is not zero.