The Cybernetic Skills Framework

An agent directed at work in this project fails in a number of recurring ways. It asserts facts about the domain that are plausible and unverified. It applies a procedure that is correct in general and wrong here, because the project's procedure differs from the one its training establishes. It cannot apply conventions it has no way of knowing, the document type system among them. And presented with a large undifferentiated catalogue of available operations, it selects poorly among them.

The four descriptions above are of surface behaviour, and a classification by surface would yield four unrelated problems. The failures identified so far are classified here instead by regulatory deficiency, meaning what the regulator lacks in the terms of the results reviewed above. The principle is adopted because the remedy follows from the deficiency rather than from the appearance, and because failures that appear unalike may require the same remedy.

On this criterion the failures fall into three classes.

Two of those four descriptions are the same failure. An agent that asserts an unverified domain fact and an agent that cannot apply a convention we invented have both acted without a fact they needed.

They differ in why the fact was missing. The domain fact exists in a knowledge document and was not loaded. The convention could not have been known at all, because this project invented it.

That difference changes the form of the remedy and not its kind. Both require the fact to be supplied. The first requires an obligation to load what already exists. The second requires the convention to be written down before there is anything to load.

The three classes organise what follows. The first two are answered by classes of document, derived under Deriving the Functions below and set out in detail in Regulatory Functions; the obligation the first imposes is treated in Grounding. The third is answered by no document, since supplying further documents is what produced it, and is treated under Variety Attenuation.

The classification is drawn from observation of practice in this project. It is not the product of a systematic study, and the three classes are not claimed to be exhaustive. That it is provisional is not a defect to be repaired before the framework can proceed. Identifying failures is itself a regulatory function, and Ultrastability specifies it as one.

What the four have in common is that none is a deficit of competence. The agent can execute the correct procedure once it is stated. It can hold a domain fact once the fact is supplied. It can select correctly from a smaller set. The capability is present and its application is unreliable.

The distinction is not merely descriptive. A capability deficit admits only one remedy, a more capable agent, and that remedy is unavailable to the framework. A regulation deficit admits a second: an apparatus that constrains the agent's conduct without altering its competence. The framework is an apparatus of the second kind, and this document specifies it as such.

Cybernetics supplies the results this document relies upon, and Cybernetics states them in general form. Three are used here.

Wiener (1948) characterises regulation by a structure common to every substrate in which it occurs: a goal, a sensor that measures the present state, and a feedback path that acts to reduce the difference. The derivation in Deriving the Functions rests on this decomposition.

Ashby (1956) establishes the law of requisite variety, that only variety can absorb variety: a regulator must command at least the variety of the disturbances it is required to counter. The law bounds what a regulator must supply, and its converse bounds what it need not.

Conant and Ashby (1970) prove that a regulator which is both maximally successful and maximally simple must be isomorphic to the system it regulates. The theorem is stated in their title: every good regulator of a system must be a model of that system. The simplicity clause carries as much weight here as the success clause, since it is what excludes modelling anything not regulated.

The theorem applies at every level of recursion, which is why model appears in this document with two referents. The framework regulates the agent's conduct, and its model is the document type system. The agent regulates the work, and its model is the domain knowledge it holds. Both follow from the same theorem applied one level apart, and each use below names which is meant.

Beer applied this apparatus to organisations in the Viable System Model, whose recursive five-system architecture this project adopts as a lens rather than as a blueprint (Cybernetic Levels).

The literature on agent skills divides on what a skill is made of. One lineage stores executable routines; the other stores verbal lessons, and the SKILL.md format belongs to the second. Skills in the Compass Skill Framework reviews both and treats the specification the format follows.

One result from that literature bears directly on the argument here. Liu et al. (2026) describe their apparatus not as an abstraction over the agent but as control upon it: a harness that wraps the base policy and intervenes on its proposed actions. Their vocabulary is worth noting, since a paper concerning programs that govern other programs employs no metalevel language at all. They characterise the purpose of such an apparatus in terms this document adopts. It exists to "surface and stabilize useful strategies that the model can often execute but does not apply reliably".

That is the distinction drawn in Regulation Rather Than Capability, reached independently and stated in the same terms. It is offered as corroboration that the framing is not idiosyncratic to this project.

By Conant and Ashby, a regulator must model what it regulates. By the simplicity clause of the same theorem, a regulator modelling what it does not regulate is not maximally simple, and is therefore not a good regulator in their sense.

The framework must model exactly what it regulates. The constraint runs in both directions, and it is the second that bounds the framework's scope.

By Ashby's law the framework must supply variety sufficient to absorb the disturbances it counters. It is not required to supply variety the agent already commands.

An agent trained on general software engineering commands most of the variety that software development presents. It knows what a compiler does, what a version control system does, and what the standard tools of the trade do. That variety is absorbed already, and supplying it again adds nothing.

The residue is what the agent does not absorb: conventions invented here, and procedures where this project departs from general practice. That residue is what the framework regulates.

Cybernetics defines the black box as the practice of characterising a system by its inputs and outputs alone, without knowledge of its internal structure, and states the condition under which the practice is appropriate: when the internals are inaccessible, or irrelevant to the regulatory task.

The second condition applies to every tool the framework invokes and does not regulate. A compiler is characterised by what it is given and what it returns. The framework does not model it. This is not an omission awaiting repair but the correct treatment of a system lying outside the regulatory task.

The three results combine into a single rule.

The framework models exactly what it regulates. It regulates where the unaided agent would not act correctly on for this project. Everything else is a black box, invoked but not modelled.

The rule settles cases that an appeal to the kind of thing under consideration does not.

The rule makes no appeal to whether a subject is a file, a directory, or a thing of some other kind. This is why a branch is modelled although it is not a file, and a compiler is not modelled although an agent uses one constantly. The criterion is regulatory relevance and nothing besides.

A regulator discharges a finite set of functions, and Wiener's decomposition names the primitives: a goal against which the state is compared, a sensor by which the state is known, and a path by which action is taken. Ashby's law adds a fourth requirement, that the regulator command sufficient variety, and Conant and Ashby a fifth, that it hold a model of what it regulates.

The classes into which this framework sorts its documents correspond to these primitives. Each class discharges one regulatory function, so the classification states a function rather than a subject matter or a form.

Declaration and recording are the two terms of a single comparison. A declaration states what ought to be true; a recording states what was in fact reached. The difference between them is the error, and reducing it is what the remaining functions exist to do.

That pairing accounts for a property the catalogue asserts without deriving it. A declaration must be checkable, because a goal an outcome cannot be compared against yields no error, and a regulator with no error signal does not regulate.

These classes are the regulatory functions (Regulatory Functions). A regulatory function denotes a regulatory function and is answerable to that account.

Three properties follow that a chosen taxonomy could not supply.

The classes are motivated. Loading exists because a regulator must hold a model of what it regulates, which is the theorem, and Grounding treats the obligation this places on an agent before it acts. Convening exists because a regulator whose variety is insufficient must acquire more, and an independent judgement is variety the agent does not itself command.

The classes are bounded. A candidate class must name a regulatory function, so the question whether some further class exists admits an answer rather than remaining a matter of preference.

And one case is settled that a classification by relation leaves open. Acting upon a thing, whether by editing, moving or deleting it, is not a distinct function. It is what an effector does, and execution is already the effector. The universal ability to operate upon an artefact is therefore absent from the list for a reason, rather than by the observation that a universal property fails to discriminate.

The strongest evidence that these are functions rather than categories of document is that one set of them serves every level of recursion. A declaration fixes a goal whether that goal is a task's acceptance criterion or a version's objective. A recording registers state whether the state is a task's status or a sprint's outcome. A taxonomy of documents would need one set per level, because a task and a version are different kinds of document. A set of functions does not, because a regulator at any horizon must still set a goal and sense the state it reached.

Everything specified so far regulates against disturbances already identified. Such a regulator is fixed. It answers the failures its regulatory functions were built for and is defeated by any other, since a disturbance outside its repertoire meets no response at all.

The framework cannot be fixed in this sense. The agents it regulates change with each model release, the project's practice changes as the system grows, and neither the failures nor their distribution are stable.

Ashby (1952) treats this case directly. An adaptive system requires a second and slower loop around the first: one that does not adjust the regulated variables but alters the structure of the regulator, and that acts only when the first loop fails to hold an essential variable within its bounds. He named the property ultrastability, and Cybernetics states it in general form.

The distinction from homeostasis is what makes the property necessary here. Homeostasis maintains a variable against disturbances the regulator was built for. Ultrastability reorganises the regulator when maintenance fails, which is the only response available to a disturbance it was not built for.

The essential variable here is whether the agent's conduct remains within acceptable bounds for the work. A failure is the signal that it has not. The second loop converts that signal into a change of structure, and the framework's structure is its regulatory functions, its document type system and its catalogue.

The second loop has three steps, and each is a regulatory function rather than an administrative one.

Detection. A departure is observed. This is measurement, and it uses the recording function: the state reached is registered against the state intended.

Classification. The departure is assigned to one of the classes above, or shown to fit none of them. The classification is therefore a live instrument and not a preamble, and a departure fitting no class is the more informative outcome, since it indicates the regulator's model is incomplete.

Conversion. The classification determines the form of the remedy, because each class names the remedy required. An absent model yields a knowledge document or a memory. A misapplied default yields a principle or a method. An excess of variety yields attenuation rather than a further document.

The loop runs in both directions. The same mechanism that adds regulation on a new failure removes it when a failure ceases to occur, which is the criterion treated under Retirement below.

The framework states the condition of this loop rather than its intended condition.

The loop is largely manual. Detection depends on a person noticing a departure during a session. Classification is informal. Conversion happens when that person directs it, and the routing of what generalises into a memory, a principle or a change of skill is performed at a session's close by request rather than by mechanism.

A loop of this kind is weak in a specific and predictable way: it registers departures a person happens to observe and is blind to those they do not. It therefore satisfies the form of ultrastability without supplying its reliability, and no claim is made here that it does. Strengthening it requires the measurement absent under Limitations, since a loop that cannot detect cannot reorganise.

The framework has two parts with different stability properties, and conflating them has obscured both.

The document type system is the model, in the sense the theorem gives the word. It states what the framework regulates. It is stable, because what requires regulation does not change when the agent changes.

The catalogue is the regulation. It is the variety the framework supplies against the disturbances presently observed. It is not stable, and it should not be.

A regulator is defined relative to its disturbances. Should the regulated system improve, the variety the regulator must supply falls.

A skill therefore becomes removable when the agent ceases to act incorrectly at that point without it. Without such a criterion the pressure runs in one direction only: skills accumulate, and nothing prompts their removal. The framework should expect to contract as the agents it regulates improve, and should treat that contraction as the apparatus working rather than as capability lost.