The Necessary Tangle · research edition 0.40 · 7 November 2026
Which distinctions change the decision?
A classification can remove detail and still support every required action. Another grouping can retain more probability-weighted information while losing the distinctions that the task needs.
Three existing profiles, two primary texts, and one original finite model. This research edition awaits human editorial acceptance.
Specify the states, probabilities, and task
These four states are invented. At the default, A has probability 97%, and B, C, and D each have 1%. The task requires the same ‘routine’ action for A and B, a callback for C, and partner coordination for D. These are labels in a finite example, with no claim about a real service.
The three share this probability equally; A receives the remainder.
Changing the task can make a previously sufficient grouping fail.
| State | Probability | Required action | Poor grouping | Default-task grouping |
|---|
Input state count4
Log₂ of input count2 bits
Input entropy H(X)
| Grouping | Output labels | Log₂ count | Entropy H(Y) | Remaining state uncertainty H(X|Y) | Task preserved? | Best expected correctness |
|---|
The poor and default-task groupings are alternative coarsenings of the full-state map. Neither refines the other. Their entropy comparison therefore does not describe a further compression from one to the other.
Best expected correctness weights the stipulated probabilities and assigns the same cost to every wrong action. It is not equal weighting of the four states or a measured algorithm accuracy. When actions tie, this example chooses the first encountered in state order A, B, C, D. At the default, ‘other’ is therefore decoded as ‘routine’, missing C and D; other equally optimal tie choices miss different states. Unequal consequences require a separate loss model.
The functional test
g = h ∘ f exactly when f(x₁) = f(x₂) implies g(x₁) = g(x₂).
Here f is the classification, g is the required action, and h is a rule that receives only the classification. Check every pair in each label group. If one group contains incompatible required actions, no such rule can be right for every member. This is a statement about the specified task; it does not decide which task ought to be chosen.
This criterion, the invented example, and its interpretation are new mathematical/editorial exposition. The primary sources below support the stated information measures and accounting distinctions.
Account for every input
A fixed stateless rule turns R, with equally probable values 0 and 1, into three repeated symbols: 000 or 111. There are still two possible output patterns. Add a separate three-valued input S, and stipulate that every pairing is possible.
Distinct output patterns
Log₂ of pattern count
Information about original R1 bit
The joint count can increase while the information about R stays at one bit. The example demonstrates input accounting; it does not demonstrate useful regulatory capacity. That would require a specified disturbance, target, effective responses, and a working arrangement. Ashby’s staged regulatory construction is a further argument, not a consequence of adding an independent input to this toy model.
Three bounded profiles
Each profile adds a primary clarification and a distinction tested by the finite model. Historical and empirical claims outside those passages remain open.
Primary evidence and its limits
Printed page numbers and one-based PDF page numbers differ in the Ashby scan. Formula signs were checked against rendered source pages. The linked primary texts retain their authorship and terms.
No new Beer management passage or empirical organisational effect is established here. No historical influence or new graph edge is inferred. Human editorial acceptance remains pending.