W. Ross Ashby gave the thing its name: variety. In cybernetics, variety is the total number of distinguishable states a system can occupy. This is not an analogy for degrees of freedom — it is the same idea restated in the language of information and control rather than mechanics. Both measure the size of the space of possible behaviour. A system with ten distinguishable states has ten degrees of freedom, and variety of ten. The vocabulary differs; the concept does not.
From this follows Ashby’s Law of Requisite Variety: only variety can absorb variety. A controller must have at least as much variety — as many degrees of freedom — as the system it is attempting to regulate. Where the controller has less variety than the system, it cannot govern the whole. It can only see an impoverished version of it. The parts it cannot see continue on regardless, doing what they do, unobserved and unmanaged.
Stafford Beer made this operational. In his work on the Viable System Model, Beer identified two basic moves: variety attenuation — reducing the variety reaching a controller, by filtering, aggregating, simplifying — and variety amplification — increasing the variety available to respond. Every management and governance structure is, in effect, an arrangement of these two mechanisms. Hierarchies are primarily attenuation devices. They compress the complexity of what is happening below into something a level above can handle. The organisation becomes legible. Decisions become possible. This is the design intent.
The problem is that attenuation does not reduce the variety of the system being managed. It reduces the variety of the model. The system retains all its degrees of freedom. The complexity that was filtered out continues to operate, invisible to the controller, exactly as before. When the real system eventually diverges from the attenuated model — as it always will — the divergence appears as unpredictable failure. It is not. The system was always that complex. The model never was.
This is the structural logic underneath what overoptimisation describes. When an organisation narrows to the measurable and sheds the rest, it is not simplifying the system — it is simplifying its model of the system while the actual system keeps all its degrees of freedom intact. The unmeasured, unrewarded, unmonitored capacity does not disappear; it waits. And what-gets-removed-does-not-come-back argues precisely this: the shed capacity eventually proves to have been load-bearing, but by the time the model catches up, the capacity is gone and the system has moved on without it.
Beer’s practical response to this was autonomy. Every subsystem in a viable organisation must be granted sufficient autonomy — sufficient retained variety — to self-regulate within its own environment. Strip that autonomy and the subsystem cannot respond to the complexity it actually faces. It is being managed through the attenuated model rather than through its own situated awareness. The freedom is still present in the situation. It is simply not being used.
The philosophical point worth sitting with is this: the variety of a living system is always already there. It is not created by good design or granted by enlightened management. It exists in the situation itself, prior to any decision about what to measure or manage. What we call control is always attenuation — a projection of that fullness into a lower-dimensional representation. The degrees of freedom are never actually removed. They are only unrecognised, and therefore unused, until they reassert themselves as the gap between what the model predicted and what the system did.
This does not make attenuation wrong. Some compression is unavoidable — no controller can hold the full variety of a complex system in view at once. The question Beer kept returning to is whether the attenuation has been designed deliberately and self-consciously, or has simply accumulated over time as a residue of what was easy to measure. An organisation that has never examined its own variety-engineering tends to have attenuated far more than necessary, and amplified far less. It mistakes the legible model for the actual system, and is governing something that no longer quite exists, while the actual system does what it will.
Related
- overoptimisation — the practical form of variety attenuation in organisations: systems narrow to the measurable and shed the unmeasured load-bearing remainder
- the-legibility-trade — a domain-specific case of variety attenuation, where ecological complexity is compressed into tradeable units at cost to systemic resilience
- three-mechanisms — overoptimisation, proxy capture, and separated knowledge are the three mechanisms by which variety is systematically attenuated at the organisational level