The Necessary Tangle · research edition 0.41
A returned signal tells us about a relation. To judge regulation, we also need to know what can act, what is protected, and which outcomes count as acceptable.
Two existing profiles · 7 November 2026 · human editorial review pending
| Question | What has to be specified? |
|---|---|
| Is there feedback? | Which parts affect which others, what returns, and what definition is being used? |
| What is its sign? | The local relation or loop, the reference condition, and the relevant operating or frequency conditions. |
| Is the response stable? | The complete dynamics and the precise stability question; a correction sign alone is insufficient. |
| Is anything usefully regulated? | Essential variables, acceptable states, disturbances, the available responses, and evidence of what actually happens. |
These distinctions organise this inquiry. They do not force all three historical sources into one vocabulary or make this finite example a general theory of organisations.
Maxwell separates decaying from growing disturbances, including oscillatory cases. He describes governor adjustment becoming too strong for stable motion. His continuous mechanical analysis does not establish the thresholds of the discrete example below.
pp271–272; PDF2–3
The authors acknowledge a broader return-of-output usage, then restrict feedback to error-controlled negative feedback. Their overshoot account says inadequate damping can make feedback positive at particular frequencies. Preserve that qualification.
pp19–20; PDF3–4
Ashby uses feedback for mutual effects, discusses a narrower return-connection usage, and warns that circuit accounts alone do not determine the behaviour of richly interconnected wholes.
§4/11, pp53–54; PDF31–32
Ashby specifies essential variables and acceptable states; protection may be passive or active. He distinguishes response to a threatened disturbance from error-based correction. His acceptable set is given outside the analysis; useful partial regulation can permit small errors.
§§10/4–10/7, pp197–201; §§12/1,12/3–12/6, pp219,221–224
These are short paraphrases. No full paper, book scan, or source-page image is included. Ashby’s rendered scan loses some arrow and set glyphs; the reading above rests on prose and extracted text, not a claimed visual reconstruction of those symbols. The 1943 paper’s clinical analogy is historical and supplies no current medical conclusion here.
In this original discrete example, for k>0, each nonzero correction opposes the current displacement from a fixed target. Change the gain and compare the analytic behaviour with the observed tolerance band. The calculation is stipulated; it has no measured service or organisational parameters.
Open the interactive response model
| Gain k | Exact result from e[n]=(1−k)ⁿe[0] |
|---|---|
| 0 | Error stays unchanged; no effective correction. |
| 0<k<1 | Error magnitude decays without changing sign. |
| 1 | Error reaches exact zero in one step. |
| 1<k<2 | Error alternates and decays. |
| 2 | Bounded undamped alternation; stable under small perturbations, but not asymptotically stable. |
| 2<k≤3 | Alternating growth in magnitude. |
With k=0.5, initial error 1, and tolerance ±0.25, step 12 lies inside the band, but the run did not stay inside it from step 0. With k=2, initial error 0.25, and tolerance ±0.25, every shown state remains inside, although the error never decays. Starting exactly at zero keeps this model there even for k=3; that run cannot establish stability under perturbation.
Fixed target: e[n+1]=(1−k)e[n]
Varying target: e[n+1]=(1−k)e[n]+r[n+1]−r[n]
The correction’s derivative with respect to the current state is −k. The multiplier of successive errors is 1−k. Confusing those two signs hides the very behaviour the example exposes. The model contains an integrating state update, with no additional delay, plant dynamics, noise, saturation, or continuing disturbance. A varying target changes the recurrence.
Maxwell’s mechanical governors have continuous dynamics; the unit-step gain thresholds here are not his result. Nor does Ashby’s formulation-specific argument about perfect error-controlled regulation forbid every exact one-step correction in a different model. Our k=1 example has a fixed target, exact measurement, no subsequent disturbance, and the stipulated update.
These are new inquiry prompts, not a validated diagnostic instrument or evidence that importing a control diagram improves a public service.
Feedback is not merely receiving comments. It is a circular relation in which consequences return to influence what happens next. The loop may counteract change, reinforce it, oscillate, delay response or fail because no actor can use the returned information. A correction sign, a stable response, and acceptable performance are separate questions. Historical authors also used different boundaries for the term.
Retained canonical definition: A circular causal or informational relation in which an output or consequence affects later input, state or action. The sign, delay, reference condition and mechanism must be stated.
Regulation is the work of keeping outcomes within acceptable ranges despite disturbance. It may operate through feedback, anticipation, constraints, buffering, redundancy, local autonomy or changes to the environment. Specify the essential variables, acceptable set, and disturbance conditions. Ashby distinguishes passive protection, response to a threatened disturbance, and error-based correction; an error need not imply that useful regulation has failed.
Retained canonical definition: A relation among disturbances, a reguland, a regulator and acceptable outcomes in which outcome variety is constrained relative to a stated goal or essential-variable range.
This edition keeps both canonical definitions, every prior evidence link, and all existing open checks. It adds four located-evidence records. It adds no graph edge or claim of historical influence. Inspect the exact proposals · Inspect the fixed baseline.
Nyquist/Bode, system dynamics, learning traditions, Beer’s organisational channels, institutional/legal regulation, and distributed cases remain open. This pass does not re-audit Conant and Ashby’s good-regulator theorem or the earlier variety packet. The mathematical example supplies a counterexample under stated assumptions, not empirical transfer to an organisation.