Why distinctions matter

Imagine a machine with a simple job: send circles to one tray and squares to another.

Here is a teal circle. Here is a teal square. You can see the difference. But the machine's sensor reports only colour. For both cards, it receives exactly the same message: teal.

How can it choose the right tray?

With this information alone, it cannot always get the answer right. It could guess. It could always choose the circle tray. But the two cards need different responses, and nothing in its input tells it which response to choose.

That is why we ask what an observer can tell apart. The answer sets a limit on what it can reliably do, including when a decision needs another observation.

This brings us to a question about mind: how could a system represent the limits of its own view, and use that knowledge?

The website makes this problem small enough to inspect completely.

There are four possibilities: a teal circle, a teal square, a gold circle, and a gold square. Their two digits are just labels for colour and shape.

Here, observer means a system with access to specified tests. We are not assuming that it has an inner experience, or imagining a little person inside the machine.

With colour alone, both teal cards give one answer, and both gold cards give another. Four different possibilities produce only two distinguishable records.

Now enable shape as well. Each card produces a different pair of answers. A rule can use the shape answer to select the correct tray.

The cards have stayed the same. What changed is the information available for choosing an action. The demonstration isolates that change so we can understand it precisely.

There is a second lesson hiding here.

Colour alone makes two groups. Shape alone also makes two groups. But only the shape groups support perfect sorting by shape.

Change the task to sorting by colour, and the useful distinction changes too.

So the number of distinctions is not enough. We need to ask what those distinctions allow the system to predict, decide, or control.

For the shape task, treating the two circles as equivalent is perfectly adequate. Their colours differ, but that difference does not affect the required answer. This gives us a small model of how a category can serve a purpose: it preserves a relevant difference while setting another aside.

That does not explain human concepts in full. It gives a proposed explanation something concrete to specify: which differences are preserved, which are ignored, and why those choices matter for the task.

We will return to this problem when a robot must tell whether a dark image comes from its surroundings or from its own camera. First, we need to be precise about what its evidence establishes.

The first linked page asks a different question: what can the available evidence establish?

We now change the rules. The website's switches gave us complete yes-or-no answers. Here, a test supplies a positive confirmation: a certificate. It tells the observer which claim the evidence supports. The observer receives these signals, rather than our view of the card. We can certify gold, or certify square.

Watch this gold card. We can see its colour, but the observer has no confirmation yet. Gold has not been established for it. That does not mean gold has been ruled out.

The gold confirmation covers two cards. The square confirmation covers two. Combine them with “and,” and their overlap identifies the gold square. Combine them with “or,” and they cover every card except the teal circle.

Together with the empty region and the whole set, these regions form a small topology. Any OR combination, and any finite AND combination, stays in this collection.

Why collect these regions? They let us check which claims the evidence can support, and compare how that changes when we add confirmations. Gold and square can jointly certify one particular card. But nothing made from these positive confirmations alone singles out the teal circle.

This is why the choice of evidence matters. A complete negative answer and the absence of a positive confirmation are different resources. The mathematics keeps that difference explicit.

The second linked page studies changes to the observation structure.

Return to the complete yes-or-no records. Keep every old test with its answers unchanged, and add a new one. States that already gave different answers remain distinguishable. A shared group may split if the new test separates its members. Repeating a test need not split anything.

That is different from a change in the world. A teal circle might become a teal square while a colour-only observer continues to receive the same answer.

One change concerns the state. The other concerns the observer's access to distinctions between states.

The page distinguishes several kinds of change to an observational representation. Adding a test is one case. Explaining how an agent finds a useful test requires a further model.

How does this connect to mind and awareness?

Think about the difference between choosing an answer and recognising that you may have answered badly. A theory of mind may need to explain both the use of information and a system's assessment of its own performance.

Research on metacognition studies that second question. For example, experiments can measure how well a person's confidence distinguishes their correct decisions from their errors, separately from how well they perform the original task.

The cards address an earlier, narrower issue: which differences are available in the specified observations at all? They do not yet show memory, attention, confidence, or a model of the observer itself.

And there is the question of experience: whether there is anything it feels like to be that system. Successful sorting, or successful self-monitoring, would not by itself establish an answer to that question.

Keeping these questions distinct lets us state what an explanation actually explains. It gives The MIND a way to formulate specific problems without treating a simple demonstration as a complete account of awareness.

Imagine a robot trying to identify a shape on a table. In our toy design, its camera returns the same black frame when it is on in a dark room and when it is off in a lit room.

The same image can therefore have two different causes. One concerns the environment. The other concerns the robot's own condition.

Now give it a reliable record of whether its camera is on. In these two cases, it can distinguish the causes and choose an appropriate next step.

We have returned to the card problem, but one of the relevant properties now belongs to the observing system itself.

A proposed self-model would represent how the robot's own sensing process affects what it can find out. It would make predictions that we can check.

In our lit-room case, the model predicts that switching its camera on will make the shape distinguishable. We can then compare that prediction with what happens.

Switching the camera on changes the robot's state. Holding the possible scenes fixed, we can also compare what it could distinguish with its camera off and with it on. The card mathematics describes that comparison; the action moves the robot between those conditions.

Knowing one's own limitations now has a concrete target: predicting when observations are insufficient for a task, and what could improve them.

But an extra input and an effective self-model are different things. Giving one system camera status while hiding it from another would demonstrate the value of that information. It would not establish that a particular self-model is better.

To test that further claim, we would compare specified models with the same available observations and comparable resources. We would vary room lighting and camera status independently, and test predictions after changing the camera's state, including cases outside their training examples.

Can the explicit model better predict whether the next image will let it identify the shape? Does its confidence distinguish correct identifications from errors? Does it request another observation when that improves identification enough to justify the cost?

It might succeed. It might add no benefit over a simpler model. Either outcome would teach us something about the proposal.

This is a possible next step for The MIND's reflective-self-model direction. It is a research target, not a reported result of the current manuscripts.

The point of the example is that a difference in the world does not automatically become a difference a system can use.

Once we make that gap explicit, we can ask better questions. Which distinctions support the task? What evidence makes them available? What changes when a new test arrives? And can a system model the limits of its own access?

B-Theory's tools make the available distinctions and supported claims explicit, so we can compare observation conditions precisely. The MIND programme proposes further models of how a system uses and changes those conditions, including its own sensing process. Connecting those models to actual minds requires further theory and evidence.

The robot leaves us with two questions to explore. What can its specified evidence establish? Follow “From observations to topology.” How does the structure of its distinctions change when another test becomes available? Follow “What changes when a new test is added.” The MIND asks what further models and experiments could turn these tools into an account of self-monitoring.

Then the four cards become a starting point for a larger investigation: what a system can distinguish, what it can do with those distinctions, and what it can find out about its own way of observing.