communicoupling
You have never once observed a thought that was not yours. Everyone you know — spouse, colleague, child — is known through a record and nothing else: what you did, what they did back. The question of what such a record can establish, and what it can never establish, has an exact answer.
W. Ross Ashby worked that answer out as the epistemics of the black box. Given a system whose interior is inaccessible, all knowledge of it derives from the protocol — the sequence of input–output pairs — and while a protocol can earn prediction and even control, it never uniquely fixes the mechanism inside. Knowledge of the box grows fastest by experiment — chosen inputs beat passive watching — converges on behaviour, and stops at behaviour's edge. Prediction is earnable; essence is not.
Below, a hidden machine. Probe it with A, B and C; its answers return as coloured glyphs; and the instrument keeps the one ledger life never shows you — a live count of how many distinct machines remain consistent with everything you have seen. Watch it fall from half a million toward its floor. The floor is almost never 1.
The protocol and what it licenses
The instrument hides a Moore machine: N internal states, each showing one of three glyphs; each input moves state to state. You never see the state — only the glyph the new state happens to show. Ashby's discipline is to admit that the record of exchanges, the protocol, is the only substance knowledge of the box can have — so the page keeps its books accordingly. With N = 3 there are 531,441 possible machines (start state fixed); every probe strikes out the ones that would have answered differently, and the counter shows how many remain. Up to three states the count is a true enumeration; at four and five, where the space swells to 7.4×1012, it is estimated from a 60,000-machine sample and shown as an estimate or a bound.
Mind where the counter stops. At absolute best, a protocol pins the mechanism up to renaming of the parts you cannot see — a floor of 2 for a fully probed three-state box, and 1 only in the special charm of two states. Every state–input cell your history never exercised leaves whole bundles of machines standing; every corrupted reading widens the plausibility band like √L, so under noise the crowd stops shrinking at all. Prediction, meanwhile, comes cheap: the page predicts each next glyph with its best surviving hypothesis (or the protocol's own frequency table when the space outruns enumeration), and accuracy climbs long before the crowd thins. The systematic prober carries the chapter's practical lesson — it asks the question the surviving models answer differently. The random prober watches whatever happens; the systematic one runs experiments, and its curves dominate at every budget.
What to try
Press The quick study. The counter starts at 576 and systematic probing drives it to 1 in about a dozen probes, coverage at 100%. Then Open the box: structurally identical. Complete knowledge of another does happen — at N = 2.
Press Surveillance vs conversation and let the random run spend its 60 probes; note the counter. Set strategy to systematic and Restart: at the same L the gold curve sits far lower. Equal budget, unequal knowledge.
Press The mimic, let it run, then Open the box. The best-fit machine agrees with every recorded observation and the verdict still reads wired differently — and, here, no experiment could ever separate them. Predictable, and unknown.
Other minds
A long marriage is a long protocol — tens of thousands of exchanges, a prediction accuracy most instruments would envy. The counter is still not at 1. What the record has earned is behaviour: you know what a raised eyebrow returns, which jokes land, what silence after a certain question means. The hypotheses remaining are the precise, countable form of the ways they could still surprise you. A job interview is a systematic probe on a brutal budget — forty minutes to collapse a candidate space; a first date runs the same instrument at a gentler setting; and both work exactly as well as their questions are ones the askers could not already answer. The surveillance preset states a hard truth about watching: the jealous reader of a partner's messages is running random inputs at enormous protocol length and a pitiful information rate, when one chosen question — asked where the hypotheses disagree — would cut the crowd in half.
The deep one is the friend of twenty easy years who, in one crisis, answers from a state you have never seen. The corner was always in the machine; the protocol had simply never run the corridor that reaches it. The mimic is the colleague you can predict perfectly and do not know at all — indistinguishable from your model of them in every act, wired otherwise inside. Noise is every misread tone, every tired answer taken as a verdict; past a certain rate of corruption, no budget of attention restores certainty. And the reveal button is the one mercy the page grants that life does not. No marriage, however long, ends with the box open. You close the lid on your model, not the mechanism, having known — richly, usefully, faithfully — only the record.
The box next door
The model the protocol earns is exactly the material the good regulator theorem demands: to steer another, you must carry an image of them, and probing is how the image gets built. Whether two boxes hold matching images can itself only be tested through the channel — say the model back and watch what returns, which is teachback. And the moment you notice that the box is also probing you — that your questions are its data — you have stepped into second-order observation, watching the watcher watch.
The mapping
| In the model | In the world |
|---|---|
| the box | The other person — mechanism inaccessible, behaviour public. |
| the protocol | Your shared history: everything either of you has ever done in the other's view. |
| probing A / B / C | Questions, jokes, tests, proposals — inputs chosen rather than merely awaited. |
| the inferred table | Your working model of them: what has tended to follow what. |
| hypotheses remaining | The ways they could still surprise you, counted. |
| open the box | The reveal life never grants; here it costs one click. |
Where it tears
Ashby's box is stationary; people are not. They learn from being probed, and the probe changes the machine — ask a person the same question often enough and they develop an answer for you. Every long relationship is two boxes rewiring each other mid-experiment, which is the observer-effect seam this page shares with reflexivity. The counter assumes the mechanism holds still. Nobody does.
The finite-state formalism radically understates a person, and the page's claim is epistemic, not ontological — it is about what a record can license, never a suggestion that persons are automata. The simplification flatters your position, if anything: enlarging the machine only swells the space of hypotheses, so every underdetermination shown here is a lower bound on the real one.
Ashby's prober reads outputs cleanly; the noise dial only smudges them. But in social life the outputs are themselves communications, often composed for the prober — the box knows what you are testing for and answers the test. That is the country of signalling games: boxes probing boxes, each protocol partly authored by the machine it is supposed to expose.